{ "schema_version": 1, "status": "prepared_not_published", "dataset_id": "yhsure/BEND2", "dataset_data_revision": "6e200d32d700d3f425173b6479ce98bd44a2d8df", "source_repository_commit": "b5d14f36b0f6b63e56d934e1f7991b1e77e25e9b", "source_files": [ { "path": "data/benchmark.json", "sha256": "f0d3657727953cf55f47c038099d75ed861b019cfd87d22314786d6f712a8b9d", "source_url": "https://github.com/MachineLearningLifeScience/BEND2/blob/b5d14f36b0f6b63e56d934e1f7991b1e77e25e9b/website/data/benchmark.json" }, { "path": "data/evaluation-updates.json", "sha256": "02a6a380db3093b8a6d75faafc094e65539cecc4ef604cd55dbefa89609c347b", "source_url": "https://github.com/MachineLearningLifeScience/BEND2/blob/b5d14f36b0f6b63e56d934e1f7991b1e77e25e9b/website/data/evaluation-updates.json" }, { "path": "provenance/table-1-source-records.csv", "sha256": "332d38d328592db946c336c0260d62c4aa02c9be005636e29bea5054fa743c28", "source_url": "https://github.com/MachineLearningLifeScience/BEND2/blob/b5d14f36b0f6b63e56d934e1f7991b1e77e25e9b/website/provenance/table-1-source-records.csv" }, { "path": "provenance/table-1.tex", "sha256": "75febce7f238262b435731ba8258f7f647111a6edf83f430f2099ada3eb5c471", "source_url": "https://github.com/MachineLearningLifeScience/BEND2/blob/b5d14f36b0f6b63e56d934e1f7991b1e77e25e9b/website/provenance/table-1.tex" } ], "precision_policy": "HF candidate values preserve sourceValue in the units already normalized by the audited website source; no second scaling, rounding, sign flip or combined index. Website display values and full measurement metadata remain alongside each result.", "protocols": [ { "id": "prior_loss_bits_per_base", "scientific_task_id": "prior", "scientific_task_label": "Prior loss", "readout_source": "shared_sequence_prior", "metric": "Hidden-base conditional log loss", "unit": "bits/base", "lower_is_better": true, "description": "Predict the hidden central base of 512 human reference-genome windows. Lower loss means better calibrated conditional probabilities.", "cohort": "512 windows of 1,200 bp from separate 1-Mb genomic blocks; central base hidden.", "cautions": [ "Interface-sensitive: NTv3 uses 12% additional masking and averages the individual losses of two queries; GENERATORv2 uses temperature-scaled probabilities.", "Only rows with a comparable nucleotide-probability interface report this task. Missing prior scores are omitted from the combined score, never imputed." ] }, { "id": "opensplice_embeddings_spearman_rho", "scientific_task_id": "opensplice", "scientific_task_label": "Splicing effects", "readout_source": "embeddings", "metric": "Macro-gene Spearman correlation", "unit": "ρ", "lower_is_better": false, "description": "Predict mutation-induced changes in alternative splicing; correlate within each held-out gene, then average genes equally.", "cohort": "590,104 interventions in 466 genes; five folds hold out complete genes for fitted readouts.", "cautions": [ "Native splice heads retain historical supervision; fitted readouts are trained on the benchmark’s source folds.", "OpenSplice is the full gene-held-out endpoint, not a cross-assay transfer score." ] }, { "id": "opensplice_native_outputs_spearman_rho", "scientific_task_id": "opensplice", "scientific_task_label": "Splicing effects", "readout_source": "native_outputs", "metric": "Macro-gene Spearman correlation", "unit": "ρ", "lower_is_better": false, "description": "Predict mutation-induced changes in alternative splicing; correlate within each held-out gene, then average genes equally.", "cohort": "590,104 interventions in 466 genes; five folds hold out complete genes for fitted readouts.", "cautions": [ "Native splice heads retain historical supervision; fitted readouts are trained on the benchmark’s source folds.", "OpenSplice is the full gene-held-out endpoint, not a cross-assay transfer score." ] }, { "id": "context_embeddings_spearman_rho", "scientific_task_id": "context", "scientific_task_label": "Cell contrast", "readout_source": "embeddings", "metric": "Macro-element Spearman correlation", "unit": "ρ", "lower_is_better": false, "description": "Predict the HepG2-minus-K562 difference in a mutation’s reporter effect, then correlate within regulatory elements and average elements equally.", "cohort": "170,292 substitutions across 128 elements; complete elements and related allelic backgrounds remain together during fitting and evaluation.", "cautions": [ "The experimental reference has physical counts for 127 of 128 elements and uses repeat subsets; model values use the pooled posterior effects." ] }, { "id": "context_native_outputs_spearman_rho", "scientific_task_id": "context", "scientific_task_label": "Cell contrast", "readout_source": "native_outputs", "metric": "Macro-element Spearman correlation", "unit": "ρ", "lower_is_better": false, "description": "Predict the HepG2-minus-K562 difference in a mutation’s reporter effect, then correlate within regulatory elements and average elements equally.", "cohort": "170,292 substitutions across 128 elements; complete elements and related allelic backgrounds remain together during fitting and evaluation.", "cautions": [ "The experimental reference has physical counts for 127 of 128 elements and uses repeat subsets; model values use the pooled posterior effects." ] }, { "id": "tissue_identification_embeddings_accuracy_percent", "scientific_task_id": "tissue_identification", "scientific_task_label": "Tissue identification", "readout_source": "embeddings", "metric": "Top-1 tissue identification accuracy", "unit": "%", "lower_is_better": false, "description": "Match a predicted expression profile across held-out genes to the correct measured tissue among 51 candidates.", "cohort": "5,782 held-out genes and 51 represented tissues; 12,146 training and 1,350 validation genes.", "cautions": [ "This holds out genes within represented tissues; it does not test transfer to unseen tissues.", "The 99.0% experimental reference averages both donor-half directions; no one-direction interval is transferred to it." ] }, { "id": "tissue_identification_native_outputs_accuracy_percent", "scientific_task_id": "tissue_identification", "scientific_task_label": "Tissue identification", "readout_source": "native_outputs", "metric": "Top-1 tissue identification accuracy", "unit": "%", "lower_is_better": false, "description": "Match a predicted expression profile across held-out genes to the correct measured tissue among 51 candidates.", "cohort": "5,782 held-out genes and 51 represented tissues; 12,146 training and 1,350 validation genes.", "cautions": [ "This holds out genes within represented tissues; it does not test transfer to unseen tissues.", "The 99.0% experimental reference averages both donor-half directions; no one-direction interval is transferred to it." ] }, { "id": "background_embeddings_spearman_rho", "scientific_task_id": "background", "scientific_task_label": "Background interaction", "readout_source": "embeddings", "metric": "Macro-element Spearman correlation", "unit": "ρ", "lower_is_better": false, "description": "Predict how the same mutation’s effect differs across two parental genetic backgrounds, isolating background dependence.", "cohort": "68,026 paired mutations in 114 elements; paired backgrounds remain together.", "cautions": [ "The current experimental reference is HepG2–K562 transfer (0.421), not the superseded replicate-subset background value.", "Transfer across cells is an experimental comparison, not a noise ceiling." ] }, { "id": "background_native_outputs_spearman_rho", "scientific_task_id": "background", "scientific_task_label": "Background interaction", "readout_source": "native_outputs", "metric": "Macro-element Spearman correlation", "unit": "ρ", "lower_is_better": false, "description": "Predict how the same mutation’s effect differs across two parental genetic backgrounds, isolating background dependence.", "cohort": "68,026 paired mutations in 114 elements; paired backgrounds remain together.", "cautions": [ "The current experimental reference is HepG2–K562 transfer (0.421), not the superseded replicate-subset background value.", "Transfer across cells is an experimental comparison, not a noise ceiling." ] }, { "id": "repair_percentile_embeddings_residual_midrank_percentile", "scientific_task_id": "repair_percentile", "scientific_task_label": "Single-edit rank", "readout_source": "embeddings", "metric": "Mean measured residual midrank percentile", "unit": "percentile", "lower_is_better": true, "description": "Choose an alternative single-base edit that matches the effect of a reference edit. Score the first nomination by its measured residual’s rank among eligible candidates.", "cohort": "331 response targets in 110 elements; 578–597 eligible substitutions per element; direct reversion is excluded.", "cautions": [ "All candidate measurements are withheld until evaluation; predictions rank proximity to the model’s own predicted reference-edit effect." ] }, { "id": "repair_percentile_native_outputs_residual_midrank_percentile", "scientific_task_id": "repair_percentile", "scientific_task_label": "Single-edit rank", "readout_source": "native_outputs", "metric": "Mean measured residual midrank percentile", "unit": "percentile", "lower_is_better": true, "description": "Choose an alternative single-base edit that matches the effect of a reference edit. Score the first nomination by its measured residual’s rank among eligible candidates.", "cohort": "331 response targets in 110 elements; 578–597 eligible substitutions per element; direct reversion is excluded.", "cautions": [ "All candidate measurements are withheld until evaluation; predictions rank proximity to the model’s own predicted reference-edit effect." ] }, { "id": "cell_selective_embeddings_utility_log2", "scientific_task_id": "cell_selective", "scientific_task_label": "Cell selectivity", "readout_source": "embeddings", "metric": "Mean cell-selective utility", "unit": "log₂ utility", "lower_is_better": false, "description": "Select one substitution or no edit to increase activity in the target cell while penalizing the absolute effect in the other cell: U = target − |other|.", "cohort": "167,958 substitutions, 280 parental backgrounds and 127 loci; both cells serve as targets; loci receive equal weight.", "cautions": [ "Native outputs use held-locus, nonnegative per-cell calibration slopes; they are not zero-shot in this task.", "Model utilities use all five RNA repeats; experimental selection uses reciprocal disjoint repeat subsets.", "Original confirmatory tests concern within-model controls; a sorted point estimate is not a significance claim." ] }, { "id": "cell_selective_native_outputs_utility_log2", "scientific_task_id": "cell_selective", "scientific_task_label": "Cell selectivity", "readout_source": "native_outputs", "metric": "Mean cell-selective utility", "unit": "log₂ utility", "lower_is_better": false, "description": "Select one substitution or no edit to increase activity in the target cell while penalizing the absolute effect in the other cell: U = target − |other|.", "cohort": "167,958 substitutions, 280 parental backgrounds and 127 loci; both cells serve as targets; loci receive equal weight.", "cautions": [ "Native outputs use held-locus, nonnegative per-cell calibration slopes; they are not zero-shot in this task.", "Model utilities use all five RNA repeats; experimental selection uses reciprocal disjoint repeat subsets.", "Original confirmatory tests concern within-model controls; a sorted point estimate is not a significance claim." ] }, { "id": "atlas_profile_embeddings_gain_over_random_log2", "scientific_task_id": "atlas_profile", "scientific_task_label": "Multi-cell gain", "readout_source": "embeddings", "metric": "Mean utility gain over intent-matched random selection", "unit": "log₂ gain", "lower_is_better": false, "description": "Select ten enhancers for each multi-cell ON/OFF program. Utility rewards the weakest requested activity and penalizes the strongest unwanted activity; report gain over intent-matched random selection.", "cohort": "2,411 measured 145-nt candidates, 16 overlapping programs and 12 cell types; select 10 per request. Fit on 5,246 separate source constructs after excluding candidate families.", "cautions": [ "The native rows and Enformer fit linear heads to native functional outputs on source constructs.", "The two RNA-repeat markers are descriptive repeat measurements, not confidence intervals or independent program replications.", "Positive gain does not imply positive absolute utility; AlphaGenome fitted has 0.833 gain but −0.096 absolute utility." ] }, { "id": "atlas_profile_native_outputs_gain_over_random_log2", "scientific_task_id": "atlas_profile", "scientific_task_label": "Multi-cell gain", "readout_source": "native_outputs", "metric": "Mean utility gain over intent-matched random selection", "unit": "log₂ gain", "lower_is_better": false, "description": "Select ten enhancers for each multi-cell ON/OFF program. Utility rewards the weakest requested activity and penalizes the strongest unwanted activity; report gain over intent-matched random selection.", "cohort": "2,411 measured 145-nt candidates, 16 overlapping programs and 12 cell types; select 10 per request. Fit on 5,246 separate source constructs after excluding candidate families.", "cautions": [ "The native rows and Enformer fit linear heads to native functional outputs on source constructs.", "The two RNA-repeat markers are descriptive repeat measurements, not confidence intervals or independent program replications.", "Positive gain does not imply positive absolute utility; AlphaGenome fitted has 0.833 gain but −0.096 absolute utility." ] } ], "model_configurations": [ { "id": "evo2_20b", "label": "Evo2 20B", "shortLabel": "Evo2 20B", "family": "Evo2", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "20B", "description": "Causal genomic language model; frozen representation readouts on functional tasks.", "sortOrder": 1 }, { "id": "ceno_1b_1m", "label": "CENO 1B", "shortLabel": "CENO 1B", "family": "CENO", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "1B", "description": "Genomic sequence model; frozen 1,024-dimensional mean embeddings for fitted tasks.", "sortOrder": 2 }, { "id": "generator_v2_3b", "label": "GENERATORv2", "shortLabel": "GENERATORv2", "family": "GENERATORv2", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "3B", "description": "Eukaryote-base causal model with sixmer tokens and nucleotide-factorized predictions.", "sortOrder": 3 }, { "id": "dnabert2", "label": "DNABERT-2", "shortLabel": "DNABERT-2", "family": "DNABERT", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "117M", "description": "Masked genomic language model with BPE tokens and frozen representation readouts.", "sortOrder": 4 }, { "id": "ntv3_100m_pre", "label": "NTv3 100M pre", "shortLabel": "NTv3 100M pre", "family": "NTv3", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "100M", "description": "Sequence-pretrained NTv3 checkpoint; frozen representation readouts.", "sortOrder": 5 }, { "id": "ntv3_100m_post", "label": "NTv3 100M post", "shortLabel": "NTv3 100M post", "family": "NTv3", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "100M", "description": "Functionally post-trained NTv3 checkpoint; frozen hidden representation readouts.", "sortOrder": 6 }, { "id": "ntv3_100m_native", "label": "NTv3 100M post · native", "shortLabel": "NTv3 100M native", "family": "NTv3", "kind": "foundation", "interface": "native", "parentModelId": "ntv3_100m_post", "parameterLabel": "100M", "description": "Same post-trained checkpoint using its native functional outputs; multi-cell still fits a linear head to those outputs.", "sortOrder": 7 }, { "id": "ntv3_650m_pre", "label": "NTv3 650M pre", "shortLabel": "NTv3 650M pre", "family": "NTv3", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "650M", "description": "Sequence-pretrained NTv3 checkpoint; frozen representation readouts.", "sortOrder": 8 }, { "id": "ntv3_650m_post", "label": "NTv3 650M post", "shortLabel": "NTv3 650M post", "family": "NTv3", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": "650M", "description": "Functionally post-trained NTv3 checkpoint; frozen hidden representation readouts. Fixed 100-point anchor for the exploratory index.", "sortOrder": 9 }, { "id": "ntv3_650m_native", "label": "NTv3 650M post · native", "shortLabel": "NTv3 650M native", "family": "NTv3", "kind": "foundation", "interface": "native", "parentModelId": "ntv3_650m_post", "parameterLabel": "650M", "description": "Same post-trained checkpoint using its native functional outputs; multi-cell still fits a linear head to those outputs.", "sortOrder": 10 }, { "id": "alphagenome_fitted", "label": "AlphaGenome", "shortLabel": "AlphaGenome", "family": "AlphaGenome", "kind": "foundation", "interface": "fitted", "parentModelId": null, "parameterLabel": null, "description": "Functionally supervised sequence model using fitted linear heads on frozen hidden representations.", "sortOrder": 11 }, { "id": "alphagenome", "label": "AlphaGenome · native", "shortLabel": "AlphaGenome native", "family": "AlphaGenome", "kind": "foundation", "interface": "native", "parentModelId": "alphagenome_fitted", "parameterLabel": null, "description": "Native functional outputs; cell-selective slopes and multi-cell native-output readout are fitted on source data.", "sortOrder": 12 }, { "id": "enformer", "label": "Enformer", "shortLabel": "Enformer", "family": "Enformer", "kind": "foundation", "interface": "native", "parentModelId": null, "parameterLabel": null, "description": "Native sequence-to-function outputs; multi-cell uses a fitted linear head on native functional features.", "sortOrder": 13 } ], "checkpoint_targets": { "evo2_20b": "arcinstitute/evo2_20b", "ceno_1b_1m": "CladeTeam/CENO-1B-1m", "generator_v2_3b": "GenerTeam/GENERator-v2-eukaryote-3b-base", "dnabert2": "zhihan1996/DNABERT-2-117M", "ntv3_100m_pre": "InstaDeepAI/NTv3_100M_pre", "ntv3_100m_post": "InstaDeepAI/NTv3_100M_post", "ntv3_100m_native": "InstaDeepAI/NTv3_100M_post", "ntv3_650m_pre": "InstaDeepAI/NTv3_650M_pre", "ntv3_650m_post": "InstaDeepAI/NTv3_650M_post", "ntv3_650m_native": "InstaDeepAI/NTv3_650M_post", "alphagenome_fitted": "google/alphagenome-all-folds", "alphagenome": "google/alphagenome-all-folds", "enformer": "EleutherAI/enformer-official-rough" }, "cells": [ { "configuration_id": "evo2_20b", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "prior", "value": 1.388, "displayValue": "1.388", "status": "evaluated", "interface": "native; 1 queries; other-position mask rate 0.0", "sourceValue": 1.3876257126643874, "sourceId": "table-1", "sourceRecordId": "evo2_20b/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.2498418346050229, "high": 1.5376494604471054, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "evaluated", "value": "1.3876257126643874", "ci_low": "1.2498418346050229", "ci_high": "1.5376494604471054", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native; 1 queries; other-position mask rate 0.0", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=evo2_20b;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata_file\\\":\\\"results/prior_benchmark_evo2_20b.json\\\",\\\"panel_sha256\\\":\\\"d4a07c4bcb98c4f5a3f94a6d82a3dfd5e01b3b5baedfc5efa075fe75e3f871f1\\\",\\\"probability_interface\\\":{\\\"base_recovery\\\":\\\"full causal quartet conditional\\\",\\\"mode\\\":\\\"causal\\\",\\\"orientation\\\":\\\"single frozen random orientation per window; no ensemble\\\",\\\"probability_interface\\\":\\\"original native interface\\\",\\\"suffix_definition\\\":\\\"log p(last 600 bp | first 600 bp), full vocabulary\\\"},\\\"protocol_sha256\\\":\\\"bd8152655f24a60932bd7082b75e6abc3fc855922c1ac15569a2a9e2f1ee2bf2\\\",\\\"source_design\\\":{\\\"batch_size\\\":1,\\\"device\\\":\\\"NVIDIA GH200 480GB\\\",\\\"evo2_version\\\":\\\"0.6.0\\\",\\\"flash_attn_version\\\":\\\"2.8.0.post2\\\",\\\"inference_dtype\\\":\\\"bfloat16 checkpoint with FP32 score difference\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"model\\\":\\\"arcinstitute/evo2_20b\\\",\\\"model_name\\\":\\\"evo2_20b\\\",\\\"parameter_count\\\":19738780160,\\\"reverse_complement_ensemble\\\":true,\\\"revision\\\":\\\"8b0f0a9a70c66367ed181a17d049b95699a28fed\\\",\\\"score\\\":\\\"bidirectional alternate-minus-reference causal logit at mutation\\\",\\\"seed\\\":20260822,\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"transformer_engine_version\\\":\\\"2.18.0\\\",\\\"unsafe_pickle_extra_state_enabled\\\":true,\\\"weight_format\\\":\\\"pinned PyTorch checkpoint\\\"}}\", \"ci_high\": \"1.5376494604471054\", \"ci_low\": \"1.2498418346050229\", \"fitting\": \"No assay labels; Markov fitted to separate genomic blocks\", \"higher_is_better\": \"False\", \"holdout\": \"512 fresh 1-Mb blocks; 1,200-bp windows; central base hidden\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"Evo2 20B\", \"library\": \"\", \"main_readout\": \"True\", \"main_table_column\": \"prior\", \"metric\": \"bits\", \"model\": \"evo2_20b\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"512\", \"notes\": \"\", \"overlay_protocol\": \"native\", \"overlay_selector\": \"figure=prior;panel=A;key=evo2_20b\", \"overlay_source\": \"reports/manuscript/figure-data/bend2-prior-points.csv\", \"overlay_source_sha256\": \"1937d554c04436e2d3a5888e4414ec66fab230047ee04c164b76aa9b21973ad7\", \"policy\": \"\", \"query\": \"\", \"readout\": \"native; 1 queries; other-position mask rate 0.0\", \"sampling_unit\": \"genomic block\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"key=evo2_20b\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/prior_benchmark.summary.csv\", \"source_sha256\": \"94de3cb8d3474c1f92e89a4decc6c4b32d1d0458e059e51cd7a266a4be1a412c\", \"status\": \"evaluated\", \"task\": \"prior\", \"task_label\": \"Hidden-base conditional log loss\", \"value\": \"1.3876257126643874\", \"variant\": \"native\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 1.3876257126643874 }, { "configuration_id": "evo2_20b", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "opensplice", "value": 0.367, "displayValue": "0.367", "status": "evaluated", "interface": "frozen-representation fitted readout", "sourceValue": 0.3667282401028527, "sourceId": "table-1", "sourceRecordId": "evo2_20b/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.35215691419807943, "high": 0.3812480803014129, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.3667282401028527", "ci_low": "0.35215691419807943", "ci_high": "0.3812480803014129", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "frozen-representation fitted readout", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=evo2_20b", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"evo2_20b\", \"label\": \"Evo2 20B\", \"value\": \"0.3667282401028527\", \"low\": \"0.35215691419807943\", \"high\": \"0.3812480803014129\", \"x\": \"\", \"role\": \"blue\", \"random\": \"False\", \"interface\": \"frozen-representation fitted readout\", \"population\": \"590104 records / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"reports/manuscript/figure-data/model-integration-20260917/evo2-splicing/opensplice_evo2_20b_gene_cv.json\", \"selector\": \"macro_spearman\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.3667282401028527 }, { "configuration_id": "evo2_20b", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "context", "value": 0.041, "displayValue": "0.041", "status": "evaluated", "interface": "Frozen representation + fitted contextual ridge", "sourceValue": 0.0408859101489466, "sourceId": "table-1", "sourceRecordId": "evo2_20b/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.0219322345988192, "high": 0.0603709624235975, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.0408859101489466", "ci_low": "0.0219322345988192", "ci_high": "0.0603709624235975", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Frozen representation + fitted contextual ridge", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=evo2_20b;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"alpha\\\":10.0,\\\"embedding_metadata_sha256\\\":\\\"551c47637f9111579ee2104093a3e673d2934e8dcbbc0700e51f581caa7348f4\\\",\\\"embedding_sha256\\\":\\\"020e03cc2a4ba3d872f70b8a804de5f2cf4fb3db29d33075b4b9d99190cb45b3\\\",\\\"folds\\\":5,\\\"projection_dim\\\":512,\\\"source_sha256\\\":\\\"b6934194264297254b98c9cc117136a502f147903598c338366f882a862ad3a3\\\",\\\"split\\\":\\\"complete held-out elements\\\",\\\"target\\\":\\\"HepG2-minus-K562 effect\\\"}\", \"ci_high\": \"0.0603709624235975\", \"ci_low\": \"0.0219322345988192\", \"fitting\": \"Five-fold element-held-out fitted readout, except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"Evo2 20B\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"evo2_20b\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Frozen representation + fitted contextual ridge\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"endpoint=context; model=evo2_20b\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_alphagenome_native.summary.csv\", \"source_sha256\": \"d7fed17feee353fef13ad62fdf8690ef1867b8fb70c2c1b4b408fd4f4ef47339\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.0408859101489466\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.0408859101489466 }, { "configuration_id": "evo2_20b", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "tissue_identification", "value": 21.6, "displayValue": "21.6", "status": "evaluated", "interface": "Embedding-fitted linear head", "sourceValue": 21.568627450980394, "sourceId": "evo2-tissue-identification", "sourceRecordId": "evo2_20b/tissue_identification", "notes": [ "Completed native-FP8 extraction with the original 131,072-bp context and 512-dimensional features; continuation used two H100 GPUs with the documented block-sharding and state adapter.", "Correctly identifies 11 of 51 tissue profiles across 5,782 held-out genes. Training-only standardization and ridge fitting, with alpha 1 selected on validation genes, were independently reproduced.", "All 19,282 feature rows are complete and all 4,489 original rows are unchanged. The shared cohort and its previously documented eight pseudoautosomal duplicate windows are unchanged." ], "uncertainty": null, "manuscriptStatus": "unevaluated", "extensionId": "coverage-20261027", "evaluatedOn": "2026-10-31", "evaluationSourceId": "evo2-tissue-identification" }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "unevaluated", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "", "source_sha256": "", "selector": "", "source_record": "", "note": "" }, "native_submission": "candidate", "native_value": 21.568627450980394 }, { "configuration_id": "evo2_20b", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "background", "value": 0.05, "displayValue": "0.050", "status": "evaluated", "interface": "Fitted out-of-fold effect differences", "sourceValue": 0.05002827879530649, "sourceId": "table-1", "sourceRecordId": "evo2_20b/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.034167087225010453, "high": 0.06576466007365578, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.05002827879530649", "ci_low": "0.034167087225010453", "ci_high": "0.06576466007365578", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Fitted out-of-fold effect differences", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=evo2_20b;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"path\\\":\\\"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_evo2_20b_srht2048_element_cv.predictions.parquet\\\",\\\"run\\\":{\\\"embedding_metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_evo2_20b.npz\\\",\\\"input_dim\\\":8192,\\\"input_metadata\\\":{\\\"batch_size\\\":1,\\\"checkpoint_file\\\":\\\"evo2_20b.pt\\\",\\\"device\\\":\\\"NVIDIA GH200 480GB\\\",\\\"embedding_dim\\\":8192,\\\"evo2_version\\\":\\\"0.6.0\\\",\\\"flash_attn_version\\\":\\\"2.8.0.post2\\\",\\\"inference_dtype\\\":\\\"FP8 input projections with BF16 remainder\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"layer\\\":\\\"blocks.20.mlp.l3\\\",\\\"model\\\":\\\"arcinstitute/evo2_20b\\\",\\\"model_name\\\":\\\"evo2_20b\\\",\\\"n_sequences\\\":169912,\\\"parameter_count\\\":19738780160,\\\"pooling\\\":\\\"mean_all_sequence_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"8b0f0a9a70c66367ed181a17d049b95699a28fed\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"sequence_lengths\\\":[189,192,197,200],\\\"storage_dtype\\\":\\\"float16\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"transformer_engine_version\\\":\\\"2.18.0\\\",\\\"vtx_version\\\":\\\"1.1.0\\\"},\\\"input_sha256\\\":\\\"626e90240d3f31ae3c80e32c7473b96bdcfd781fbbe8a7b0efbd0d303cf3ab6e\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":2048,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"63e6b4af059061f96548f9b4a1ea2a953a71c50f1a0ed52f002b3106d844743e\\\",\\\"sign_vector_sha256\\\":\\\"bfe568cf3823c426e886be0513b1df97fd6f37f314fd473bc6ae28a1c7699a56\\\",\\\"storage_dtype\\\":\\\"float16\\\"},\\\"embedding_path\\\":\\\"data/cache/embeddings/siraj_evo2_20b_srht2048.npz\\\",\\\"folds\\\":5,\\\"group_column\\\":\\\"unit_id\\\",\\\"ridge_alpha\\\":10.0,\\\"target\\\":\\\"delta_y\\\"},\\\"sha256\\\":\\\"01772b2d44fb489098259abedc7da0efc0bd240af3cc7d7d18ff7d692500d652\\\"}\", \"ci_high\": \"0.06576466007365578\", \"ci_low\": \"0.034167087225010453\", \"fitting\": \"Out-of-fold effects with complete elements held out; except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Same mutation on two backgrounds; 114 elements / 68,026 pairs\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"Evo2 20B\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"evo2_20b\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Fitted out-of-fold effect differences\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=evo2_20b\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_background_interaction.summary.csv\", \"source_sha256\": \"13fc40cd51049436fb416d7c5c5de74d9c4b5f084be6534276d64ab3e1698c4d\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj sequence-background interaction\", \"value\": \"0.05002827879530649\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.05002827879530649 }, { "configuration_id": "evo2_20b", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "repair_percentile", "value": 43.5, "displayValue": "43.5", "status": "evaluated", "interface": "held-element response readout", "sourceValue": 43.5447413260353, "sourceId": "table-1", "sourceRecordId": "evo2_20b/repair_percentile", "notes": [ "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" ], "uncertainty": { "level": 0.95, "low": 39.5343568427543, "high": 47.47643784175484, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "43.5447413260353", "ci_low": "39.5343568427543", "ci_high": "47.47643784175484", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "held-element response readout", "source_file": "source/repair-summary.csv", "source_sha256": "433befcbc4b8c0d5d7b3779c2a2aa1c15257bd5a4cd28aaeb35ae0ccbd1ca5a7", "selector": "model=evo2_20b;value=100*candidate_percentile", "source_record": "{\"candidate_percentile\": \"0.435447413260353\", \"ci_high\": \"0.47476437841754837\", \"ci_low\": \"0.395343568427543\", \"elements\": \"110\", \"family\": \"foundation\", \"label\": \"Evo2 20B\", \"model\": \"evo2_20b\", \"resolved_instances\": \"331\", \"top1_success_ci_high\": \"0.16363636363636364\", \"top1_success_ci_low\": \"0.07878787878787878\", \"top1_success_rate\": \"0.11893939393939393\"}", "note": "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" }, "native_submission": "candidate", "native_value": 43.5447413260353 }, { "configuration_id": "evo2_20b", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "cell_selective", "value": -0.35, "displayValue": "-0.350", "status": "evaluated", "interface": "fitted common+contrast, whole-locus OOF", "sourceValue": -0.3500543707285904, "sourceId": "table-1", "sourceRecordId": "evo2_20b/cell_selective", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": -0.4183947346639236, "high": -0.2855335601651783, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "-0.3500543707285904", "ci_low": "-0.4183947346639236", "ci_high": "-0.2855335601651783", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted common+contrast, whole-locus OOF", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=evo2_20b;task=cell_selective", "source_record": "{\"ci_high\": \"-0.2855335601651783\", \"ci_low\": \"-0.4183947346639236\", \"interface\": \"fitted common+contrast, whole-locus OOF\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"evo2_20b\", \"n_groups\": \"127\", \"n_units\": \"280\", \"result_file\": \"output/missing-cellselective-evaluations-20260922/cached-four/table_ready_rows.csv\", \"result_sha256\": \"dd125a82b3ff26b8c9bcc9318aeb8562f84226b3248993627fc369a6850fc58c\", \"source_path\": \"/Users/fvk220/Documents/New project/output/missing-cellselective-evaluations-20260922/cached-four/summary.csv\", \"source_sha256\": \"220905a9d94b369a0350a642248012db11264ea09ccd88edc122c34a9be4b902\", \"task\": \"cell_selective\", \"value\": \"-0.3500543707285904\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": -0.3500543707285904 }, { "configuration_id": "evo2_20b", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "arcinstitute/evo2_20b", "measurement": { "modelId": "evo2_20b", "taskId": "atlas_profile", "value": 0.275, "displayValue": "0.275", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 0.2745767447929687, "sourceId": "table-1", "sourceRecordId": "evo2_20b/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "evo2_20b", "label": "Evo2 20B", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.2745767447929687", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/evo2/summary.csv", "source_sha256": "51c8b34292dc05d79888c3b3301a49fcaf327b17e17189d68f4092f0bf3b5ff2", "selector": "model=Evo2 20B;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"Evo2 20B\", \"repeat1\": \"0.2326410245026707\", \"repeat2\": \"0.3165124650832666\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/evo2/summary.csv\", \"source_sha256\": \"51c8b34292dc05d79888c3b3301a49fcaf327b17e17189d68f4092f0bf3b5ff2\", \"status\": \"independently replayed selection and utility\", \"value\": \"0.2745767447929687\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.2745767447929687 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "prior", "value": 1.549, "displayValue": "1.549", "status": "evaluated", "interface": "native; 1 queries; other-position mask rate 0.0", "sourceValue": 1.548726503070408, "sourceId": "table-1", "sourceRecordId": "ceno_1b_1m/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.3848975957560896, "high": 1.7360729767331837, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "evaluated", "value": "1.548726503070408", "ci_low": "1.3848975957560896", "ci_high": "1.7360729767331837", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native; 1 queries; other-position mask rate 0.0", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ceno_1b_1m;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata_file\\\":\\\"results/prior_benchmark_ceno_1b_1m.json\\\",\\\"panel_sha256\\\":\\\"d4a07c4bcb98c4f5a3f94a6d82a3dfd5e01b3b5baedfc5efa075fe75e3f871f1\\\",\\\"probability_interface\\\":{\\\"base_recovery\\\":\\\"full causal quartet conditional\\\",\\\"mode\\\":\\\"causal\\\",\\\"orientation\\\":\\\"single frozen random orientation per window; no ensemble\\\",\\\"probability_interface\\\":\\\"original native interface\\\",\\\"suffix_definition\\\":\\\"log p(last 600 bp | first 600 bp), full vocabulary\\\"},\\\"protocol_sha256\\\":\\\"bd8152655f24a60932bd7082b75e6abc3fc855922c1ac15569a2a9e2f1ee2bf2\\\",\\\"source_design\\\":{\\\"batch_size\\\":1024,\\\"dependencies\\\":[\\\"CENO_repository@1a5f1685a52c226452b3d1b2c5be27f6da70f111\\\"],\\\"device\\\":\\\"NVIDIA GH200 480GB\\\",\\\"inference_dtype\\\":\\\"bfloat16\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"CladeTeam/CENO-1B-1m\\\",\\\"n_unique_sequences\\\":339824,\\\"parameter_count\\\":1302429280,\\\"random_init\\\":false,\\\"reverse_complement_ensemble\\\":true,\\\"revision\\\":\\\"fa8089cb4c9e22665190a2a7514a6346ec37d95d\\\",\\\"score\\\":\\\"alternate-minus-reference summed causal log likelihood\\\",\\\"seed\\\":20260822,\\\"tokenizer_class\\\":\\\"CENOCharLevelTokenizer\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"transformers_version\\\":\\\"4.57.6\\\",\\\"weight_format\\\":\\\"safetensors\\\"}}\", \"ci_high\": \"1.7360729767331837\", \"ci_low\": \"1.3848975957560896\", \"fitting\": \"No assay labels; Markov fitted to separate genomic blocks\", \"higher_is_better\": \"False\", \"holdout\": \"512 fresh 1-Mb blocks; 1,200-bp windows; central base hidden\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"CENO 1B\", \"library\": \"\", \"main_readout\": \"True\", \"main_table_column\": \"prior\", \"metric\": \"bits\", \"model\": \"ceno_1b_1m\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"512\", \"notes\": \"\", \"overlay_protocol\": \"native\", \"overlay_selector\": \"figure=prior;panel=A;key=ceno_1b_1m\", \"overlay_source\": \"reports/manuscript/figure-data/bend2-prior-points.csv\", \"overlay_source_sha256\": \"1937d554c04436e2d3a5888e4414ec66fab230047ee04c164b76aa9b21973ad7\", \"policy\": \"\", \"query\": \"\", \"readout\": \"native; 1 queries; other-position mask rate 0.0\", \"sampling_unit\": \"genomic block\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"key=ceno_1b_1m\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/prior_benchmark.summary.csv\", \"source_sha256\": \"94de3cb8d3474c1f92e89a4decc6c4b32d1d0458e059e51cd7a266a4be1a412c\", \"status\": \"evaluated\", \"task\": \"prior\", \"task_label\": \"Hidden-base conditional log loss\", \"value\": \"1.548726503070408\", \"variant\": \"native\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 1.548726503070408 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "opensplice", "value": 0.295, "displayValue": "0.295", "status": "evaluated", "interface": "Embedding-fitted linear head", "sourceValue": 0.29544350315143425, "sourceId": "ceno-full-opensplice", "sourceRecordId": "ceno_1b_1m/opensplice", "notes": [ "All embeddings were recomputed uniformly in FP32; the published gene splits and readout settings are unchanged.", "The table retains the published solver tolerance (1e-4). A converged solution gives 0.303 and is archived as a separate numerical check." ], "uncertainty": { "level": 0.95, "low": 0.28159746286866255, "high": 0.30892280281213297, "type": "pointwise percentile bootstrap interval", "samplingUnit": "gene", "sourceId": "ceno-full-opensplice", "replicates": 2000, "seed": 20260822 }, "manuscriptStatus": "unevaluated", "extensionId": "coverage-20261027", "evaluatedOn": "2026-10-27", "evaluationSourceId": "ceno-full-opensplice" }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "unevaluated", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "", "source_sha256": "", "selector": "", "source_record": "", "note": "No matching full OpenSplice gene-held-out result available; transfer scores not substituted." }, "native_submission": "candidate", "native_value": 0.29544350315143425 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "context", "value": 0.036, "displayValue": "0.036", "status": "evaluated", "interface": "frozen 1024-dimensional mean embeddings + paired-response ridge", "sourceValue": 0.0356386997088576, "sourceId": "table-1", "sourceRecordId": "ceno_1b_1m/context", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": { "level": 0.95, "low": 0.0182474419071388, "high": 0.0529378199353923, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.0356386997088576", "ci_low": "0.0182474419071388", "ci_high": "0.0529378199353923", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "frozen 1024-dimensional mean embeddings + paired-response ridge", "source_file": "provenance/table-completion-20260923/sources/table-gap-ceno-20260923/table-ready.csv", "source_sha256": "4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006", "selector": "model=CENO 1B;endpoint=cell_contrast", "source_record": "{\"endpoint\": \"cell_contrast\", \"high\": \"0.0529378199353923\", \"interface\": \"frozen 1024-dimensional mean embeddings + paired-response ridge\", \"low\": \"0.0182474419071388\", \"model\": \"CENO 1B\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-gap-ceno-20260923/table-ready.csv\", \"source_sha256\": \"4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006\", \"status\": \"independently audited\", \"value\": \"0.0356386997088576\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 0.0356386997088576 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "tissue_identification", "value": 9.8, "displayValue": "9.8", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 9.80392156862745, "sourceId": "table-1", "sourceRecordId": "ceno_1b_1m/tissue_identification", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": null }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "9.80392156862745", "ci_low": "5.88235294117647", "ci_high": "19.6078431372549", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/gtex-ceno-results/results.csv", "source_sha256": "0bb4d35a729fa8e0532bdf7e4d6d7d4038f19deab655ed7e41be722b82f687aa", "selector": "model=CENO 1B;endpoint=tissue_id", "source_record": "{\"endpoint\": \"tissue_id\", \"high\": \"19.6078431372549\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"5.88235294117647\", \"model\": \"CENO 1B\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/gtex-ceno-results/results.csv\", \"source_sha256\": \"0bb4d35a729fa8e0532bdf7e4d6d7d4038f19deab655ed7e41be722b82f687aa\", \"status\": \"evaluated; see independent audit\", \"value\": \"9.80392156862745\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 9.80392156862745 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "background", "value": 0.01, "displayValue": "0.010", "status": "evaluated", "interface": "fitted mean-effect frozen readout; paired backgrounds", "sourceValue": 0.0100908337947297, "sourceId": "table-1", "sourceRecordId": "ceno_1b_1m/background", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": { "level": 0.95, "low": -0.0073462810463968, "high": 0.0272147551454775, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.0100908337947297", "ci_low": "-0.0073462810463968", "ci_high": "0.0272147551454775", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted mean-effect frozen readout; paired backgrounds", "source_file": "provenance/table-completion-20260923/sources/table-gap-ceno-20260923/table-ready.csv", "source_sha256": "4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006", "selector": "model=CENO 1B;endpoint=background_interaction", "source_record": "{\"endpoint\": \"background_interaction\", \"high\": \"0.0272147551454775\", \"interface\": \"fitted mean-effect frozen readout; paired backgrounds\", \"low\": \"-0.0073462810463968\", \"model\": \"CENO 1B\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-gap-ceno-20260923/table-ready.csv\", \"source_sha256\": \"4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006\", \"status\": \"independently audited\", \"value\": \"0.0100908337947297\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 0.0100908337947297 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "repair_percentile", "value": 48.9, "displayValue": "48.9", "status": "evaluated", "interface": "fitted frozen readout; whole-locus OOF; rank fraction (multiply by100 for table)", "sourceValue": 48.8559753317775, "sourceId": "table-1", "sourceRecordId": "ceno_1b_1m/repair_percentile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": { "level": 0.95, "low": 44.872684835068846, "high": 52.69490833117898, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "48.8559753317775", "ci_low": "44.872684835068846", "ci_high": "52.69490833117898", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted frozen readout; whole-locus OOF; rank fraction (multiply by100 for table)", "source_file": "provenance/table-completion-20260923/sources/table-gap-ceno-20260923/table-ready.csv", "source_sha256": "4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006", "selector": "model=CENO 1B;endpoint=single_edit_rank", "source_record": "{\"endpoint\": \"single_edit_rank\", \"high\": \"52.69490833117898\", \"interface\": \"fitted frozen readout; whole-locus OOF; rank fraction (multiply by100 for table)\", \"low\": \"44.872684835068846\", \"model\": \"CENO 1B\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-gap-ceno-20260923/table-ready.csv\", \"source_sha256\": \"4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006\", \"status\": \"independently audited\", \"value\": \"48.8559753317775\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 48.8559753317775 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "cell_selective", "value": -0.318, "displayValue": "-0.318", "status": "evaluated", "interface": "fitted common+contrast, whole-locus OOF", "sourceValue": -0.3182551000994145, "sourceId": "table-1", "sourceRecordId": "ceno_1b_1m/cell_selective", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": { "level": 0.95, "low": -0.3923549620430436, "high": -0.2484480266321445, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "-0.3182551000994145", "ci_low": "-0.3923549620430436", "ci_high": "-0.2484480266321445", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted common+contrast, whole-locus OOF", "source_file": "provenance/table-completion-20260923/sources/table-gap-ceno-20260923/table-ready.csv", "source_sha256": "4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006", "selector": "model=CENO 1B;endpoint=cell_selectivity", "source_record": "{\"endpoint\": \"cell_selectivity\", \"high\": \"-0.2484480266321445\", \"interface\": \"fitted common+contrast, whole-locus OOF\", \"low\": \"-0.3923549620430436\", \"model\": \"CENO 1B\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-gap-ceno-20260923/table-ready.csv\", \"source_sha256\": \"4e640fa9084679754d9f62d8a98d17b7085b0432e215f9654513468dd1585006\", \"status\": \"independently audited\", \"value\": \"-0.3182551000994145\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": -0.3182551000994145 }, { "configuration_id": "ceno_1b_1m", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "CladeTeam/CENO-1B-1m", "measurement": { "modelId": "ceno_1b_1m", "taskId": "atlas_profile", "value": -0.581, "displayValue": "-0.581", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": -0.581495968622555, "sourceId": "table-1", "sourceRecordId": "ceno_1b_1m/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "ceno_1b_1m", "label": "CENO 1B", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "-0.581495968622555", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/five-models/summary.csv", "source_sha256": "e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092", "selector": "model=CENO 1B;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"CENO 1B\", \"repeat1\": \"-0.58408802539808\", \"repeat2\": \"-0.57890391184703\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/five-models/summary.csv\", \"source_sha256\": \"e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092\", \"status\": \"independently replayed selection and utility\", \"value\": \"-0.581495968622555\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": -0.581495968622555 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "prior", "value": 1.711, "displayValue": "1.711", "status": "evaluated", "interface": "Temperature-scaled native conditional probabilities", "sourceValue": 1.7114884266131933, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.6518856399072959, "high": 1.774712041432719, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "evaluated", "value": "1.7114884266131933", "ci_low": "1.6518856399072959", "ci_high": "1.774712041432719", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Temperature-scaled native conditional probabilities", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=generator_v2_3b;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"calibration_panel_sha256\\\":\\\"4d20c93dd710606d79d4b51f45b41dde420ff4872aa117f2d8124ced9ad126e9\\\",\\\"evaluation_panel_sha256\\\":\\\"d4a07c4bcb98c4f5a3f94a6d82a3dfd5e01b3b5baedfc5efa075fe75e3f871f1\\\",\\\"parameters\\\":{\\\"boundary_hit\\\":null,\\\"calibration_bits_fitted\\\":1.7835494433283647,\\\"calibration_bits_native\\\":2.511828772100248,\\\"calibration_source\\\":\\\"prior_sequence_generator_v2_3b_fns\\\",\\\"calibration_windows\\\":1024,\\\"evaluation_source\\\":\\\"prior_benchmark_generator_v2_3b_fns\\\",\\\"log_temperature\\\":1.361606738218507,\\\"lower_bound\\\":0.1,\\\"temperature\\\":3.9024584963958495,\\\"upper_bound\\\":100.0},\\\"sources\\\":{\\\"prior_benchmark_generator_v2_3b_fns\\\":{\\\"metadata_sha256\\\":\\\"2c2b49e7d4c4f592cc5322a45590bc5f115047b7911a895368d6d6891a3bd2eb\\\",\\\"path\\\":\\\"results/prior_benchmark_generator_v2_3b_fns.predictions.parquet\\\",\\\"sha256\\\":\\\"36e27ddf8a0507fb25393b85873f38a6914be37096cb879bf75a4663c36e0d8f\\\"},\\\"prior_sequence_generator_v2_3b_fns\\\":{\\\"metadata_sha256\\\":\\\"b203300948a67fba9bf6a5c1a6c855e17e5679b2efa4c112d05c3067598b7c04\\\",\\\"path\\\":\\\"results/prior_sequence_generator_v2_3b_fns.predictions.parquet\\\",\\\"sha256\\\":\\\"5b4e389c4fd1a1bd43717ff7b2028e1661694980ca3c77da0f26a2731b4cccab\\\"}},\\\"temperature\\\":\\\"3.9024584963958495\\\"}\", \"ci_high\": \"1.774712041432719\", \"ci_low\": \"1.6518856399072959\", \"fitting\": \"One scalar temperature fitted on separate original 1,024 genomic windows; no assay labels\", \"higher_is_better\": \"False\", \"holdout\": \"512 fresh 1-Mb blocks; 1,200-bp windows; central base hidden\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"GENERATOR v2 3B\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"prior\", \"metric\": \"bits\", \"model\": \"generator_v2_3b\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"512\", \"notes\": \"\", \"overlay_protocol\": \"temperature calibrated on 1024 separate windows\", \"overlay_selector\": \"figure=prior;panel=A;key=generator_v2_3b\", \"overlay_source\": \"reports/manuscript/figure-data/bend2-prior-points.csv\", \"overlay_source_sha256\": \"1937d554c04436e2d3a5888e4414ec66fab230047ee04c164b76aa9b21973ad7\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Temperature-scaled native conditional probabilities\", \"sampling_unit\": \"genomic block\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model_key=generator_v2_3b_fns; readout=temperature\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/prior_calibration.summary.csv\", \"source_sha256\": \"88588953aeeaf592262890b4bc9e6443d855e01a9ba46db91edee5d0e45c42a1\", \"status\": \"evaluated\", \"task\": \"prior\", \"task_label\": \"Hidden-base conditional log loss\", \"value\": \"1.7114884266131933\", \"variant\": \"temperature calibrated\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 1.7114884266131933 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "opensplice", "value": 0.388, "displayValue": "0.388", "status": "evaluated", "interface": "gene-held-out fitted readout", "sourceValue": 0.388096180972446, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.373739563390649, "high": 0.4021376584878464, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.388096180972446", "ci_low": "0.373739563390649", "ci_high": "0.4021376584878464", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "gene-held-out fitted readout", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=generator_v2_3b", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"generator_v2_3b\", \"label\": \"GENERATORv2 3B\", \"value\": \"0.388096180972446\", \"low\": \"0.373739563390649\", \"high\": \"0.4021376584878464\", \"x\": \"\", \"role\": \"blue\", \"random\": \"False\", \"interface\": \"gene-held-out fitted readout\", \"population\": \"590,104 interventions / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"results/opensplice_model_suite.summary.csv\", \"selector\": \"model_key=generator_v2_3b;fields=pretrained_*\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.388096180972446 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "context", "value": 0.066, "displayValue": "0.066", "status": "evaluated", "interface": "Frozen representation + fitted contextual ridge", "sourceValue": 0.06642642407560108, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.050235505888559545, "high": 0.08292127442263753, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.06642642407560108", "ci_low": "0.050235505888559545", "ci_high": "0.08292127442263753", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Frozen representation + fitted contextual ridge", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=generator_v2_3b;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_generator_v2_3b_pretrained.npz\\\",\\\"input_dim\\\":3072,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"causal-lm\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[],\\\"embedding_dim\\\":3072,\\\"implementation_module\\\":\\\"transformers_modules.GenerTeam.GENERator_hyphen_v2_hyphen_eukaryote_hyphen_3b_hyphen_base.7dc01bccce5b65e15141170538afdc2ff09d8dde.modeling_generator\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"GenerTeam/GENERator-v2-eukaryote-3b-base\\\",\\\"n_sequences\\\":169912,\\\"parameter_count\\\":2998262784,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"7dc01bccce5b65e15141170538afdc2ff09d8dde\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"]},\\\"input_sha256\\\":\\\"f4924bebdc54f0e9ee57b82f3eed0313f483ef1c6f10482f4d695477eb6db026\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"8da75a652dc3cc2c0d171b597d33f6fa3b956e9fbe7050cd99d32a1a24d96118\\\",\\\"sign_vector_sha256\\\":\\\"704f64c62879c703550e985f92bc8be6865c4bb5f4665dbddb92c1c6b64d5291\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":4096,\\\"zero_padding_dimensions\\\":1024},\\\"path\\\":\\\"data/cache/embeddings/siraj_generator_v2_3b_pretrained_srht512.npz\\\",\\\"sha256\\\":\\\"2f6dfa4c03a2472656192a9e807c2f4e70d47b64d53fe906d17b7531349867f0\\\"}\", \"ci_high\": \"0.08292127442263753\", \"ci_low\": \"0.050235505888559545\", \"fitting\": \"Five-fold element-held-out fitted readout, except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"GENERATOR v2 3B\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"generator_v2_3b\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Frozen representation + fitted contextual ridge\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=generator_v2_3b_pretrained\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_contextual_suite.summary.csv\", \"source_sha256\": \"c948c9ed32febd7e045aa76df7f52cf49da13bed3ecd45e8bb497188a10f3049\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.06642642407560108\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.06642642407560108 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "tissue_identification", "value": 17.6, "displayValue": "17.6", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 17.64705882352941, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/tissue_identification", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Fixed centered 98,298-bp crop; 16,384-token limit including BOS." ], "uncertainty": null }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "17.64705882352941", "ci_low": "9.80392156862745", "ci_high": "21.56862745098039", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/gtex-extra-results/results.csv", "source_sha256": "0e135a631faea79fe6d4c34c5eda5798e0c9a4932a21f3017650ee0013be4f4d", "selector": "model=GENERATORv2;endpoint=tissue_id", "source_record": "{\"endpoint\": \"tissue_id\", \"high\": \"21.56862745098039\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"9.80392156862745\", \"model\": \"GENERATORv2\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/gtex-extra-results/results.csv\", \"source_sha256\": \"0e135a631faea79fe6d4c34c5eda5798e0c9a4932a21f3017650ee0013be4f4d\", \"status\": \"evaluated; see independent audit\", \"value\": \"17.64705882352941\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Fixed centered 98,298-bp crop; 16,384-token limit including BOS." }, "native_submission": "candidate", "native_value": 17.64705882352941 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "background", "value": 0.046, "displayValue": "0.046", "status": "evaluated", "interface": "Fitted out-of-fold effect differences", "sourceValue": 0.04578286189430206, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.03043472220980968, "high": 0.06165299685548288, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.04578286189430206", "ci_low": "0.03043472220980968", "ci_high": "0.06165299685548288", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Fitted out-of-fold effect differences", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=generator_v2_3b;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"path\\\":\\\"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_generator_v2_3b_element_cv.predictions.parquet\\\",\\\"run\\\":{\\\"embedding_metadata\\\":{\\\"architecture\\\":\\\"causal-lm\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[],\\\"embedding_dim\\\":3072,\\\"implementation_module\\\":\\\"transformers_modules.GenerTeam.GENERator_hyphen_v2_hyphen_eukaryote_hyphen_3b_hyphen_base.7dc01bccce5b65e15141170538afdc2ff09d8dde.modeling_generator\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"GenerTeam/GENERator-v2-eukaryote-3b-base\\\",\\\"n_sequences\\\":169912,\\\"parameter_count\\\":2998262784,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"7dc01bccce5b65e15141170538afdc2ff09d8dde\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"]},\\\"embedding_path\\\":\\\"data/cache/embeddings/siraj_generator_v2_3b_pretrained.npz\\\",\\\"folds\\\":5,\\\"group_column\\\":\\\"unit_id\\\",\\\"ridge_alpha\\\":10.0,\\\"target\\\":\\\"delta_y\\\"},\\\"sha256\\\":\\\"92aa825671dbd5716f2fad3aaab45216b91f0412e02740d2fce75823bb6aaa23\\\"}\", \"ci_high\": \"0.06165299685548288\", \"ci_low\": \"0.03043472220980968\", \"fitting\": \"Out-of-fold effects with complete elements held out; except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Same mutation on two backgrounds; 114 elements / 68,026 pairs\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"GENERATOR v2 3B\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"generator_v2_3b\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Fitted out-of-fold effect differences\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=generator_v2_3b\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_background_interaction.summary.csv\", \"source_sha256\": \"13fc40cd51049436fb416d7c5c5de74d9c4b5f084be6534276d64ab3e1698c4d\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj sequence-background interaction\", \"value\": \"0.04578286189430206\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.04578286189430206 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "repair_percentile", "value": 44.7, "displayValue": "44.7", "status": "evaluated", "interface": "held-element response readout", "sourceValue": 44.70248556452466, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/repair_percentile", "notes": [ "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" ], "uncertainty": { "level": 0.95, "low": 40.86696598240759, "high": 48.53537264105221, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "44.70248556452466", "ci_low": "40.86696598240759", "ci_high": "48.53537264105221", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "held-element response readout", "source_file": "source/repair-summary.csv", "source_sha256": "433befcbc4b8c0d5d7b3779c2a2aa1c15257bd5a4cd28aaeb35ae0ccbd1ca5a7", "selector": "model=generator_v2_3b;value=100*candidate_percentile", "source_record": "{\"candidate_percentile\": \"0.4470248556452466\", \"ci_high\": \"0.48535372641052216\", \"ci_low\": \"0.40866965982407594\", \"elements\": \"110\", \"family\": \"foundation\", \"label\": \"GENERATOR-v2 3B\", \"model\": \"generator_v2_3b\", \"resolved_instances\": \"331\", \"top1_success_ci_high\": \"0.11439393939393938\", \"top1_success_ci_low\": \"0.04772727272727273\", \"top1_success_rate\": \"0.07878787878787878\"}", "note": "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" }, "native_submission": "candidate", "native_value": 44.70248556452466 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "cell_selective", "value": -0.256, "displayValue": "-0.256", "status": "evaluated", "interface": "fitted common+contrast, whole-locus OOF", "sourceValue": -0.2563913486763549, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/cell_selective", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": -0.3260562618114235, "high": -0.1900645318874209, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "-0.2563913486763549", "ci_low": "-0.3260562618114235", "ci_high": "-0.1900645318874209", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted common+contrast, whole-locus OOF", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=generator_v2_3b;task=cell_selective", "source_record": "{\"ci_high\": \"-0.1900645318874209\", \"ci_low\": \"-0.3260562618114235\", \"interface\": \"fitted common+contrast, whole-locus OOF\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"generator_v2_3b\", \"n_groups\": \"127\", \"n_units\": \"280\", \"result_file\": \"output/missing-cellselective-evaluations-20260922/cached-four/table_ready_rows.csv\", \"result_sha256\": \"dd125a82b3ff26b8c9bcc9318aeb8562f84226b3248993627fc369a6850fc58c\", \"source_path\": \"/Users/fvk220/Documents/New project/output/missing-cellselective-evaluations-20260922/cached-four/summary.csv\", \"source_sha256\": \"220905a9d94b369a0350a642248012db11264ea09ccd88edc122c34a9be4b902\", \"task\": \"cell_selective\", \"value\": \"-0.2563913486763549\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": -0.2563913486763549 }, { "configuration_id": "generator_v2_3b", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "GenerTeam/GENERator-v2-eukaryote-3b-base", "measurement": { "modelId": "generator_v2_3b", "taskId": "atlas_profile", "value": 0.234, "displayValue": "0.234", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 0.2344857975720761, "sourceId": "table-1", "sourceRecordId": "generator_v2_3b/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "generator_v2_3b", "label": "GENERATORv2", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.2344857975720761", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/five-models/summary.csv", "source_sha256": "e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092", "selector": "model=GENERATORv2;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"GENERATORv2\", \"repeat1\": \"0.242512050867217\", \"repeat2\": \"0.2264595442769352\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/five-models/summary.csv\", \"source_sha256\": \"e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092\", \"status\": \"independently replayed selection and utility\", \"value\": \"0.2344857975720761\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.2344857975720761 }, { "configuration_id": "dnabert2", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "prior", "value": null, "displayValue": "—", "status": "unsupported", "interface": null, "sourceValue": null, "sourceId": "table-1", "sourceRecordId": "dnabert2/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": null, "manuscriptStatus": "unevaluated", "missingReason": "DNABERT-2 predicts multi-base tokens. A forced single-base query was tested, but its tokenization mismatch prevents a directly comparable prior score.", "extensionId": "coverage-20261027", "diagnosticSourceId": "dnabert2-prior-diagnostic" }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "unevaluated", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=dnabert2;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"\", \"ci_low\": \"\", \"fitting\": \"\", \"higher_is_better\": \"\", \"holdout\": \"\", \"initialization\": \"trained\", \"interval\": \"\", \"label\": \"dnabert2\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"prior\", \"metric\": \"\", \"model\": \"dnabert2\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"\", \"sampling_unit\": \"\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"\", \"source_sha256\": \"\", \"status\": \"unevaluated\", \"task\": \"prior\", \"task_label\": \"\", \"value\": \"\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "omitted_missing_or_unsupported_never_zero" }, { "configuration_id": "dnabert2", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "opensplice", "value": 0.198, "displayValue": "0.198", "status": "evaluated", "interface": "gene-held-out fitted readout", "sourceValue": 0.1981902308981018, "sourceId": "table-1", "sourceRecordId": "dnabert2/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.1875703889427766, "high": 0.2097891470967729, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.1981902308981018", "ci_low": "0.1875703889427766", "ci_high": "0.2097891470967729", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "gene-held-out fitted readout", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=dnabert2", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"dnabert2\", \"label\": \"DNABERT-2\", \"value\": \"0.1981902308981018\", \"low\": \"0.1875703889427766\", \"high\": \"0.2097891470967729\", \"x\": \"\", \"role\": \"blue\", \"random\": \"False\", \"interface\": \"gene-held-out fitted readout\", \"population\": \"590,104 interventions / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"results/opensplice_model_suite.summary.csv\", \"selector\": \"model_key=dnabert2;fields=pretrained_*\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.1981902308981018 }, { "configuration_id": "dnabert2", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "context", "value": 0.076, "displayValue": "0.076", "status": "evaluated", "interface": "frozen 768-dimensional mean embeddings + paired-response ridge", "sourceValue": 0.0761897168480282, "sourceId": "table-1", "sourceRecordId": "dnabert2/context", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": 0.0655674881071053, "high": 0.0866062326246656, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.0761897168480282", "ci_low": "0.0655674881071053", "ci_high": "0.0866062326246656", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "frozen 768-dimensional mean embeddings + paired-response ridge", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=dnabert2;task=context", "source_record": "{\"ci_high\": \"0.0866062326246656\", \"ci_low\": \"0.0655674881071053\", \"interface\": \"frozen 768-dimensional mean embeddings + paired-response ridge\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"dnabert2\", \"n_groups\": \"128\", \"n_units\": \"170292\", \"result_file\": \"reports/missing-siraj-evaluations-20260922/table-ready.csv\", \"result_sha256\": \"bc190435e9380e15275082f5439c05e783b2dd94f188f20201cbfea64566406f\", \"source_path\": \"/Users/fvk220/Documents/New project/reports/missing-siraj-evaluations-20260922/contextual/siraj_dnabert2_context.predictions.parquet\", \"source_sha256\": \"a73b097d9fe14de65039329000c940a2bcccfe3009b58f36b3a9f676e6a59798\", \"task\": \"context\", \"value\": \"0.0761897168480282\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": 0.0761897168480282 }, { "configuration_id": "dnabert2", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "tissue_identification", "value": 5.9, "displayValue": "5.9", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 5.88235294117647, "sourceId": "table-1", "sourceRecordId": "dnabert2/tissue_identification", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": null }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "5.88235294117647", "ci_low": "1.9607843137254901", "ci_high": "9.80392156862745", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/gtex-extra-results/results.csv", "source_sha256": "0e135a631faea79fe6d4c34c5eda5798e0c9a4932a21f3017650ee0013be4f4d", "selector": "model=DNABERT-2;endpoint=tissue_id", "source_record": "{\"endpoint\": \"tissue_id\", \"high\": \"9.80392156862745\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"1.9607843137254901\", \"model\": \"DNABERT-2\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/gtex-extra-results/results.csv\", \"source_sha256\": \"0e135a631faea79fe6d4c34c5eda5798e0c9a4932a21f3017650ee0013be4f4d\", \"status\": \"evaluated; see independent audit\", \"value\": \"5.88235294117647\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 5.88235294117647 }, { "configuration_id": "dnabert2", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "background", "value": 0.02, "displayValue": "0.020", "status": "evaluated", "interface": "frozen 768-dimensional mean embeddings + ridge", "sourceValue": 0.0197275780667494, "sourceId": "table-1", "sourceRecordId": "dnabert2/background", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": 0.0073320822009257, "high": 0.0321235628360173, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.0197275780667494", "ci_low": "0.0073320822009257", "ci_high": "0.0321235628360173", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "frozen 768-dimensional mean embeddings + ridge", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=dnabert2;task=background", "source_record": "{\"ci_high\": \"0.0321235628360173\", \"ci_low\": \"0.0073320822009257\", \"interface\": \"frozen 768-dimensional mean embeddings + ridge\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"dnabert2\", \"n_groups\": \"114\", \"n_units\": \"68026\", \"result_file\": \"reports/missing-siraj-evaluations-20260922/table-ready.csv\", \"result_sha256\": \"bc190435e9380e15275082f5439c05e783b2dd94f188f20201cbfea64566406f\", \"source_path\": \"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_dnabert2_pretrained_element_cv.predictions.parquet\", \"source_sha256\": \"7c5b993f753b96a79fc7ce4d54f0c79c7cfe400038f1e1f1f6a7c888427b2e64\", \"task\": \"background\", \"value\": \"0.0197275780667494\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": 0.0197275780667494 }, { "configuration_id": "dnabert2", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "repair_percentile", "value": 48.6, "displayValue": "48.6", "status": "evaluated", "interface": "frozen 768-dimensional mean embeddings + ridge", "sourceValue": 48.55974077076406, "sourceId": "table-1", "sourceRecordId": "dnabert2/repair_percentile", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": 44.99529018373341, "high": 52.24416780360783, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "48.55974077076406", "ci_low": "44.99529018373341", "ci_high": "52.24416780360783", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "frozen 768-dimensional mean embeddings + ridge", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=dnabert2;task=repair_percentile", "source_record": "{\"ci_high\": \"52.24416780360783\", \"ci_low\": \"44.99529018373341\", \"interface\": \"frozen 768-dimensional mean embeddings + ridge\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"dnabert2\", \"n_groups\": \"110\", \"n_units\": \"331\", \"result_file\": \"reports/missing-siraj-evaluations-20260922/table-ready.csv\", \"result_sha256\": \"bc190435e9380e15275082f5439c05e783b2dd94f188f20201cbfea64566406f\", \"source_path\": \"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_dnabert2_pretrained_element_cv.predictions.parquet\", \"source_sha256\": \"7c5b993f753b96a79fc7ce4d54f0c79c7cfe400038f1e1f1f6a7c888427b2e64\", \"task\": \"repair_percentile\", \"value\": \"48.55974077076406\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": 48.55974077076406 }, { "configuration_id": "dnabert2", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "cell_selective", "value": -0.302, "displayValue": "-0.302", "status": "evaluated", "interface": "fitted common+contrast, whole-locus OOF", "sourceValue": -0.3018458156636018, "sourceId": "table-1", "sourceRecordId": "dnabert2/cell_selective", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": -0.3774399515069673, "high": -0.2300589702969386, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "-0.3018458156636018", "ci_low": "-0.3774399515069673", "ci_high": "-0.2300589702969386", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted common+contrast, whole-locus OOF", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=dnabert2;task=cell_selective", "source_record": "{\"ci_high\": \"-0.2300589702969386\", \"ci_low\": \"-0.3774399515069673\", \"interface\": \"fitted common+contrast, whole-locus OOF\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"dnabert2\", \"n_groups\": \"127\", \"n_units\": \"280\", \"result_file\": \"output/missing-cellselective-evaluations-20260922/dnabert2/table_ready_rows.csv\", \"result_sha256\": \"337b333c7e7574d81b463eaf4d9b31be7f466d9ea44cb69a9a20dc6c3828f03b\", \"source_path\": \"/Users/fvk220/Documents/New project/output/missing-cellselective-evaluations-20260922/dnabert2/summary.csv\", \"source_sha256\": \"58ecffa000e9fdb7995860e610b1dfde5097ad29bebdd4eb58cbd9a5d5bf874b\", \"task\": \"cell_selective\", \"value\": \"-0.3018458156636018\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": -0.3018458156636018 }, { "configuration_id": "dnabert2", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "zhihan1996/DNABERT-2-117M", "measurement": { "modelId": "dnabert2", "taskId": "atlas_profile", "value": 0.233, "displayValue": "0.233", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 0.2328719532775636, "sourceId": "table-1", "sourceRecordId": "dnabert2/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "dnabert2", "label": "DNABERT-2", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.2328719532775636", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/five-models/summary.csv", "source_sha256": "e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092", "selector": "model=DNABERT-2;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"DNABERT-2\", \"repeat1\": \"0.2235294734470943\", \"repeat2\": \"0.2422144331080329\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/five-models/summary.csv\", \"source_sha256\": \"e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092\", \"status\": \"independently replayed selection and utility\", \"value\": \"0.2328719532775636\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.2328719532775636 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "prior", "value": 1.537, "displayValue": "1.537", "status": "evaluated", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "sourceValue": 1.537171017457747, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.453831449879011, "high": 1.6202512118874777, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "evaluated", "value": "1.537171017457747", "ci_low": "1.453831449879011", "ci_high": "1.6202512118874777", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_pre;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata_file\\\":\\\"results/prior_benchmark_ntv3_100m_pre_native.json\\\",\\\"panel_sha256\\\":\\\"d4a07c4bcb98c4f5a3f94a6d82a3dfd5e01b3b5baedfc5efa075fe75e3f871f1\\\",\\\"probability_interface\\\":{},\\\"protocol_sha256\\\":\\\"bd8152655f24a60932bd7082b75e6abc3fc855922c1ac15569a2a9e2f1ee2bf2\\\",\\\"source_design\\\":{\\\"architecture\\\":\\\"masked-lm\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"model\\\":\\\"InstaDeepAI/NTv3_100M_pre\\\",\\\"parameter_count\\\":106463419,\\\"random_init\\\":false,\\\"reverse_complement_ensemble\\\":true,\\\"revision\\\":\\\"5c685dca15891f5c5b80e0c930e23b87a217e441\\\",\\\"score\\\":\\\"mean forward/reverse masked alternate-minus-reference log probability\\\",\\\"seed\\\":20260822,\\\"species\\\":null,\\\"tokenization\\\":\\\"character-level, no special tokens, padded to a multiple of 128\\\"}}\", \"ci_high\": \"1.6202512118874777\", \"ci_low\": \"1.453831449879011\", \"fitting\": \"Fixed 12% mask rate from the original protocol; no fitted readout or assay labels\", \"higher_is_better\": \"False\", \"holdout\": \"512 fresh 1-Mb blocks; 1,200-bp windows; central base hidden\", \"initialization\": \"trained\", \"interval\": \"95% interval over 512 genomic blocks\", \"label\": \"NTv3 100M pre\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"prior\", \"metric\": \"bits\", \"model\": \"ntv3_100m_pre\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"512\", \"notes\": \"Mean individual-query scores, not a probability ensemble. Same 12% protocol for all NTv3 sizes/stages.\", \"overlay_protocol\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"overlay_selector\": \"figure=prior;panel=A;key=ntv3_100m_pre\", \"overlay_source\": \"reports/manuscript/figure-data/bend2-prior-points.csv\", \"overlay_source_sha256\": \"1937d554c04436e2d3a5888e4414ec66fab230047ee04c164b76aa9b21973ad7\", \"policy\": \"\", \"query\": \"One query per prediction; loss averaged over two fixed mask draws\", \"readout\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"sampling_unit\": \"genomic block\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model_key=ntv3_100m_pre;readout=mask12_single_query\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/prior_calibration.summary.csv\", \"source_sha256\": \"88588953aeeaf592262890b4bc9e6443d855e01a9ba46db91edee5d0e45c42a1\", \"status\": \"evaluated\", \"task\": \"prior\", \"task_label\": \"Hidden-base conditional log loss\", \"value\": \"1.5371710174577471\", \"variant\": \"12% single query\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 1.537171017457747 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "opensplice", "value": 0.349, "displayValue": "0.349", "status": "evaluated", "interface": "gene-held-out fitted readout", "sourceValue": 0.3490793743190476, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.334962566678737, "high": 0.3632066159269334, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.3490793743190476", "ci_low": "0.334962566678737", "ci_high": "0.3632066159269334", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "gene-held-out fitted readout", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=ntv3_100m_pre", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"ntv3_100m_pre\", \"label\": \"NTv3 100M pre\", \"value\": \"0.3490793743190476\", \"low\": \"0.334962566678737\", \"high\": \"0.3632066159269334\", \"x\": \"\", \"role\": \"blue\", \"random\": \"False\", \"interface\": \"gene-held-out fitted readout\", \"population\": \"590,104 interventions / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"results/opensplice_model_suite.summary.csv\", \"selector\": \"model_key=ntv3_100m_pre;fields=pretrained_*\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.3490793743190476 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "context", "value": 0.076, "displayValue": "0.076", "status": "evaluated", "interface": "Frozen representation + fitted contextual ridge", "sourceValue": 0.07602431417082979, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.06254040669021495, "high": 0.08948517254946839, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.07602431417082979", "ci_low": "0.06254040669021495", "ci_high": "0.08948517254946839", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Frozen representation + fitted contextual ridge", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_pre;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_ntv3_100m_pre_pretrained.npz\\\",\\\"input_dim\\\":768,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"masked-lm\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[\\\"InstaDeepAI/ntv3_base_model@0ecff3637f0d3ba5b686d1095083218157c2ca34\\\"],\\\"embedding_dim\\\":768,\\\"implementation_module\\\":\\\"transformers_modules.InstaDeepAI.ntv3_base_model.0ecff3637f0d3ba5b686d1095083218157c2ca34.modeling_ntv3_pretrained\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/element_split.parquet\\\",\\\"input_sha256\\\":\\\"64941bdb63305fb6555614ef8934f0f563758999c71d01a2ad44e393fe7ae261\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"InstaDeepAI/NTv3_100M_pre\\\",\\\"n_sequences\\\":169912,\\\"pad_to_multiple_of\\\":128,\\\"parameter_count\\\":106463419,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"5c685dca15891f5c5b80e0c930e23b87a217e441\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"species\\\":null,\\\"state_dict\\\":null,\\\"state_dict_metadata\\\":null},\\\"input_sha256\\\":\\\"861b17dc5357296e68633dc08fafa7ababcb7e7a1aef1499c2cfda54865c54d5\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"596a446b5957012b15a1d71cea44589fdb1760ad47e7e16aad2fb04fd119174d\\\",\\\"sign_vector_sha256\\\":\\\"8391a78e7212adf2f66ae412f9da37f7c1aec3bfe88f5323221fcdf6c44a736b\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":1024,\\\"zero_padding_dimensions\\\":256},\\\"path\\\":\\\"data/cache/embeddings/siraj_ntv3_100m_pre_pretrained_srht512.npz\\\",\\\"sha256\\\":\\\"374bb4d09ec38ba0730026433a56762352571576f670e9972b66bf4121909682\\\"}\", \"ci_high\": \"0.08948517254946839\", \"ci_low\": \"0.06254040669021495\", \"fitting\": \"Five-fold element-held-out fitted readout, except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"NTv3 100M pre\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_100m_pre\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Frozen representation + fitted contextual ridge\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=ntv3_100m_pre_pretrained\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_contextual_suite.summary.csv\", \"source_sha256\": \"c948c9ed32febd7e045aa76df7f52cf49da13bed3ecd45e8bb497188a10f3049\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.07602431417082979\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.07602431417082979 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "tissue_identification", "value": 7.8, "displayValue": "7.8", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 7.8431372549019605, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/tissue_identification", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "7.8431372549019605", "ci_low": "5.88235294117647", "ci_high": "15.686274509803921", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/gtex-nt-results/results.csv", "source_sha256": "7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f", "selector": "model=NTv3 100M pre;endpoint=tissue_id", "source_record": "{\"endpoint\": \"tissue_id\", \"high\": \"15.686274509803921\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"5.88235294117647\", \"model\": \"NTv3 100M pre\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/gtex-nt-results/results.csv\", \"source_sha256\": \"7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f\", \"status\": \"evaluated; see independent audit\", \"value\": \"7.8431372549019605\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 7.8431372549019605 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "background", "value": 0.01, "displayValue": "0.010", "status": "evaluated", "interface": "frozen mean embeddings + 512-dimensional SRHT + ridge", "sourceValue": 0.0101713052936839, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/background", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": -0.0003871016379826, "high": 0.0205973859159844, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.0101713052936839", "ci_low": "-0.0003871016379826", "ci_high": "0.0205973859159844", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "frozen mean embeddings + 512-dimensional SRHT + ridge", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=ntv3_100m_pre;task=background", "source_record": "{\"ci_high\": \"0.0205973859159844\", \"ci_low\": \"-0.0003871016379826\", \"interface\": \"frozen mean embeddings + 512-dimensional SRHT + ridge\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"ntv3_100m_pre\", \"n_groups\": \"114\", \"n_units\": \"68026\", \"result_file\": \"reports/missing-siraj-evaluations-20260922/table-ready.csv\", \"result_sha256\": \"bc190435e9380e15275082f5439c05e783b2dd94f188f20201cbfea64566406f\", \"source_path\": \"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_ntv3_100m_pre_pretrained_element_cv.predictions.parquet\", \"source_sha256\": \"9ca015ac9af5073334fe9002cb9563d21da4dbdcaaa7383863a17dc8b1b2a219\", \"task\": \"background\", \"value\": \"0.0101713052936839\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": 0.0101713052936839 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "repair_percentile", "value": 47.3, "displayValue": "47.3", "status": "evaluated", "interface": "frozen mean embeddings + 512-dimensional SRHT + ridge", "sourceValue": 47.26204943355816, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/repair_percentile", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": 43.77074567142887, "high": 50.80598748052813, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "47.26204943355816", "ci_low": "43.77074567142887", "ci_high": "50.80598748052813", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "frozen mean embeddings + 512-dimensional SRHT + ridge", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=ntv3_100m_pre;task=repair_percentile", "source_record": "{\"ci_high\": \"50.80598748052813\", \"ci_low\": \"43.77074567142887\", \"interface\": \"frozen mean embeddings + 512-dimensional SRHT + ridge\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"ntv3_100m_pre\", \"n_groups\": \"110\", \"n_units\": \"331\", \"result_file\": \"reports/missing-siraj-evaluations-20260922/table-ready.csv\", \"result_sha256\": \"bc190435e9380e15275082f5439c05e783b2dd94f188f20201cbfea64566406f\", \"source_path\": \"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_ntv3_100m_pre_pretrained_element_cv.predictions.parquet\", \"source_sha256\": \"9ca015ac9af5073334fe9002cb9563d21da4dbdcaaa7383863a17dc8b1b2a219\", \"task\": \"repair_percentile\", \"value\": \"47.26204943355816\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": 47.26204943355816 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "cell_selective", "value": -0.207, "displayValue": "-0.207", "status": "evaluated", "interface": "fitted common+contrast, whole-locus OOF", "sourceValue": -0.2072906378917048, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/cell_selective", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": -0.2649251232963834, "high": -0.1514757646934401, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "-0.2072906378917048", "ci_low": "-0.2649251232963834", "ci_high": "-0.1514757646934401", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted common+contrast, whole-locus OOF", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=ntv3_100m_pre;task=cell_selective", "source_record": "{\"ci_high\": \"-0.1514757646934401\", \"ci_low\": \"-0.2649251232963834\", \"interface\": \"fitted common+contrast, whole-locus OOF\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"ntv3_100m_pre\", \"n_groups\": \"127\", \"n_units\": \"280\", \"result_file\": \"output/missing-cellselective-evaluations-20260922/cached-four/table_ready_rows.csv\", \"result_sha256\": \"dd125a82b3ff26b8c9bcc9318aeb8562f84226b3248993627fc369a6850fc58c\", \"source_path\": \"/Users/fvk220/Documents/New project/output/missing-cellselective-evaluations-20260922/cached-four/summary.csv\", \"source_sha256\": \"220905a9d94b369a0350a642248012db11264ea09ccd88edc122c34a9be4b902\", \"task\": \"cell_selective\", \"value\": \"-0.2072906378917048\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": -0.2072906378917048 }, { "configuration_id": "ntv3_100m_pre", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "InstaDeepAI/NTv3_100M_pre", "measurement": { "modelId": "ntv3_100m_pre", "taskId": "atlas_profile", "value": 0.339, "displayValue": "0.339", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 0.3391168926403413, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_pre/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_100m_pre", "label": "NTv3 100M pre", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.3391168926403413", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/five-models/summary.csv", "source_sha256": "e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092", "selector": "model=NTv3 100M pre;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"NTv3 100M pre\", \"repeat1\": \"0.3315341857386868\", \"repeat2\": \"0.3466995995419958\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/five-models/summary.csv\", \"source_sha256\": \"e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092\", \"status\": \"independently replayed selection and utility\", \"value\": \"0.3391168926403413\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.3391168926403413 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "prior", "value": 1.492, "displayValue": "1.492", "status": "evaluated", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "sourceValue": 1.492461935201394, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.3927389998597406, "high": 1.5916175825667722, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "evaluated", "value": "1.492461935201394", "ci_low": "1.3927389998597406", "ci_high": "1.5916175825667722", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_post;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata_file\\\":\\\"results/prior_benchmark_ntv3_100m_post_native.json\\\",\\\"panel_sha256\\\":\\\"d4a07c4bcb98c4f5a3f94a6d82a3dfd5e01b3b5baedfc5efa075fe75e3f871f1\\\",\\\"probability_interface\\\":{},\\\"protocol_sha256\\\":\\\"bd8152655f24a60932bd7082b75e6abc3fc855922c1ac15569a2a9e2f1ee2bf2\\\",\\\"source_design\\\":{\\\"architecture\\\":\\\"base\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"model\\\":\\\"InstaDeepAI/NTv3_100M_post\\\",\\\"parameter_count\\\":120058868,\\\"random_init\\\":false,\\\"reverse_complement_ensemble\\\":true,\\\"revision\\\":\\\"b7292f0bd1b5b004d28561783a1890e1ec6f94fd\\\",\\\"score\\\":\\\"mean forward/reverse masked alternate-minus-reference log probability\\\",\\\"seed\\\":20260822,\\\"species\\\":\\\"human\\\",\\\"tokenization\\\":\\\"character-level, no special tokens, padded to a multiple of 128\\\"}}\", \"ci_high\": \"1.5916175825667722\", \"ci_low\": \"1.3927389998597406\", \"fitting\": \"Fixed 12% mask rate from the original protocol; no fitted readout or assay labels\", \"higher_is_better\": \"False\", \"holdout\": \"512 fresh 1-Mb blocks; 1,200-bp windows; central base hidden\", \"initialization\": \"trained\", \"interval\": \"95% interval over 512 genomic blocks\", \"label\": \"NTv3 100M post\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"prior\", \"metric\": \"bits\", \"model\": \"ntv3_100m_post\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"512\", \"notes\": \"Mean individual-query scores, not a probability ensemble. Same 12% protocol for all NTv3 sizes/stages.\", \"overlay_protocol\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"overlay_selector\": \"figure=prior;panel=A;key=ntv3_100m_post\", \"overlay_source\": \"reports/manuscript/figure-data/bend2-prior-points.csv\", \"overlay_source_sha256\": \"1937d554c04436e2d3a5888e4414ec66fab230047ee04c164b76aa9b21973ad7\", \"policy\": \"\", \"query\": \"One query per prediction; loss averaged over two fixed mask draws\", \"readout\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"sampling_unit\": \"genomic block\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model_key=ntv3_100m_post;readout=mask12_single_query\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/prior_calibration.summary.csv\", \"source_sha256\": \"88588953aeeaf592262890b4bc9e6443d855e01a9ba46db91edee5d0e45c42a1\", \"status\": \"evaluated\", \"task\": \"prior\", \"task_label\": \"Hidden-base conditional log loss\", \"value\": \"1.4924619352013939\", \"variant\": \"12% single query\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 1.492461935201394 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "opensplice", "value": 0.502, "displayValue": "0.502", "status": "evaluated", "interface": "gene-held-out fitted readout", "sourceValue": 0.5021724547333987, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.4877427406119896, "high": 0.5160748744142409, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.5021724547333987", "ci_low": "0.4877427406119896", "ci_high": "0.5160748744142409", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "gene-held-out fitted readout", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=ntv3_100m_post", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"ntv3_100m_post\", \"label\": \"NTv3 100M post\", \"value\": \"0.5021724547333987\", \"low\": \"0.4877427406119896\", \"high\": \"0.5160748744142409\", \"x\": \"\", \"role\": \"teal\", \"random\": \"False\", \"interface\": \"gene-held-out fitted readout\", \"population\": \"590,104 interventions / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"results/opensplice_model_suite.summary.csv\", \"selector\": \"model_key=ntv3_100m_post;fields=pretrained_*\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.5021724547333987 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "context", "value": 0.244, "displayValue": "0.244", "status": "evaluated", "interface": "Frozen representation + fitted contextual ridge", "sourceValue": 0.24415434895768506, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.21488583455283577, "high": 0.2733822431961442, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.24415434895768506", "ci_low": "0.21488583455283577", "ci_high": "0.2733822431961442", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Frozen representation + fitted contextual ridge", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_post;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_ntv3_100m_post_pretrained.npz\\\",\\\"input_dim\\\":768,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"base\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[\\\"InstaDeepAI/ntv3_base_model@0ecff3637f0d3ba5b686d1095083218157c2ca34\\\"],\\\"embedding_dim\\\":768,\\\"implementation_module\\\":\\\"transformers_modules.InstaDeepAI.ntv3_base_model.0ecff3637f0d3ba5b686d1095083218157c2ca34.modeling_ntv3_posttrained\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/element_split.parquet\\\",\\\"input_sha256\\\":\\\"64941bdb63305fb6555614ef8934f0f563758999c71d01a2ad44e393fe7ae261\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"InstaDeepAI/NTv3_100M_post\\\",\\\"n_sequences\\\":169912,\\\"pad_to_multiple_of\\\":128,\\\"parameter_count\\\":120058868,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"b7292f0bd1b5b004d28561783a1890e1ec6f94fd\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"species\\\":\\\"human\\\",\\\"state_dict\\\":null,\\\"state_dict_metadata\\\":null},\\\"input_sha256\\\":\\\"bb4dd88b7e6ff8de3b18e499b1d0cbfdc98af0dd93b6a74923861d9286c38caf\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"596a446b5957012b15a1d71cea44589fdb1760ad47e7e16aad2fb04fd119174d\\\",\\\"sign_vector_sha256\\\":\\\"8391a78e7212adf2f66ae412f9da37f7c1aec3bfe88f5323221fcdf6c44a736b\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":1024,\\\"zero_padding_dimensions\\\":256},\\\"path\\\":\\\"data/cache/embeddings/siraj_ntv3_100m_post_pretrained_srht512.npz\\\",\\\"sha256\\\":\\\"bfa3d89db62c1a39181a826e570fdff1941a96607d05b8283268170eb5132782\\\"}\", \"ci_high\": \"0.2733822431961442\", \"ci_low\": \"0.21488583455283577\", \"fitting\": \"Five-fold element-held-out fitted readout, except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"NTv3 100M post\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_100m_post\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Frozen representation + fitted contextual ridge\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=ntv3_100m_post_pretrained\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_contextual_suite.summary.csv\", \"source_sha256\": \"c948c9ed32febd7e045aa76df7f52cf49da13bed3ecd45e8bb497188a10f3049\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.24415434895768506\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.24415434895768506 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "tissue_identification", "value": 51.0, "displayValue": "51.0", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 50.98039215686274, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/tissue_identification", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "50.98039215686274", "ci_low": "35.29411764705882", "ci_high": "56.86274509803921", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/gtex-nt-results/results.csv", "source_sha256": "7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f", "selector": "model=NTv3 100M post;endpoint=tissue_id", "source_record": "{\"endpoint\": \"tissue_id\", \"high\": \"56.86274509803921\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"35.29411764705882\", \"model\": \"NTv3 100M post\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/gtex-nt-results/results.csv\", \"source_sha256\": \"7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f\", \"status\": \"evaluated; see independent audit\", \"value\": \"50.98039215686274\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 50.98039215686274 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "background", "value": 0.139, "displayValue": "0.139", "status": "evaluated", "interface": "Fitted out-of-fold effect differences", "sourceValue": 0.1390996911117981, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.11000098933347734, "high": 0.16842399530305707, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.1390996911117981", "ci_low": "0.11000098933347734", "ci_high": "0.16842399530305707", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Fitted out-of-fold effect differences", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_post;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"path\\\":\\\"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_ntv3_100m_post_pretrained_element_cv.predictions.parquet\\\",\\\"run\\\":{\\\"embedding_metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_ntv3_100m_post_pretrained.npz\\\",\\\"input_dim\\\":768,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"base\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[\\\"InstaDeepAI/ntv3_base_model@0ecff3637f0d3ba5b686d1095083218157c2ca34\\\"],\\\"embedding_dim\\\":768,\\\"implementation_module\\\":\\\"transformers_modules.InstaDeepAI.ntv3_base_model.0ecff3637f0d3ba5b686d1095083218157c2ca34.modeling_ntv3_posttrained\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/element_split.parquet\\\",\\\"input_sha256\\\":\\\"64941bdb63305fb6555614ef8934f0f563758999c71d01a2ad44e393fe7ae261\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"InstaDeepAI/NTv3_100M_post\\\",\\\"n_sequences\\\":169912,\\\"pad_to_multiple_of\\\":128,\\\"parameter_count\\\":120058868,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"b7292f0bd1b5b004d28561783a1890e1ec6f94fd\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"species\\\":\\\"human\\\",\\\"state_dict\\\":null,\\\"state_dict_metadata\\\":null},\\\"input_sha256\\\":\\\"bb4dd88b7e6ff8de3b18e499b1d0cbfdc98af0dd93b6a74923861d9286c38caf\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"596a446b5957012b15a1d71cea44589fdb1760ad47e7e16aad2fb04fd119174d\\\",\\\"sign_vector_sha256\\\":\\\"8391a78e7212adf2f66ae412f9da37f7c1aec3bfe88f5323221fcdf6c44a736b\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":1024,\\\"zero_padding_dimensions\\\":256},\\\"embedding_path\\\":\\\"data/cache/embeddings/siraj_ntv3_100m_post_pretrained_srht512.npz\\\",\\\"folds\\\":5,\\\"group_column\\\":\\\"unit_id\\\",\\\"ridge_alpha\\\":10.0,\\\"target\\\":\\\"delta_y\\\"},\\\"sha256\\\":\\\"82f7049dd337213ab0b7dfc041381af820606f25a5f35df6502da6c5be5d350e\\\"}\", \"ci_high\": \"0.16842399530305707\", \"ci_low\": \"0.11000098933347734\", \"fitting\": \"Out-of-fold effects with complete elements held out; except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Same mutation on two backgrounds; 114 elements / 68,026 pairs\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"NTv3 100M post\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_100m_post\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Fitted out-of-fold effect differences\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=ntv3_100m_post\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_background_interaction.summary.csv\", \"source_sha256\": \"13fc40cd51049436fb416d7c5c5de74d9c4b5f084be6534276d64ab3e1698c4d\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj sequence-background interaction\", \"value\": \"0.1390996911117981\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.1390996911117981 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "repair_percentile", "value": 30.3, "displayValue": "30.3", "status": "evaluated", "interface": "held-element response readout", "sourceValue": 30.26982345573195, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/repair_percentile", "notes": [ "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" ], "uncertainty": { "level": 0.95, "low": 25.935172116779516, "high": 34.77778045880539, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "30.26982345573195", "ci_low": "25.935172116779516", "ci_high": "34.77778045880539", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "held-element response readout", "source_file": "source/repair-summary.csv", "source_sha256": "433befcbc4b8c0d5d7b3779c2a2aa1c15257bd5a4cd28aaeb35ae0ccbd1ca5a7", "selector": "model=ntv3_100m_post;value=100*candidate_percentile", "source_record": "{\"candidate_percentile\": \"0.3026982345573195\", \"ci_high\": \"0.34777780458805385\", \"ci_low\": \"0.25935172116779515\", \"elements\": \"110\", \"family\": \"functional\", \"label\": \"NTv3 100M post\", \"model\": \"ntv3_100m_post\", \"resolved_instances\": \"331\", \"top1_success_ci_high\": \"0.23181818181818178\", \"top1_success_ci_low\": \"0.13484848484848486\", \"top1_success_rate\": \"0.18181818181818182\"}", "note": "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" }, "native_submission": "candidate", "native_value": 30.26982345573195 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "cell_selective", "value": 0.211, "displayValue": "0.211", "status": "evaluated", "interface": "held-locus common-effect plus cell-contrast readouts", "sourceValue": 0.21063830067077918, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/cell_selective", "notes": [ "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" ], "uncertainty": { "level": 0.95, "low": 0.12386243944096922, "high": 0.2989373349286535, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "0.21063830067077918", "ci_low": "0.12386243944096922", "ci_high": "0.2989373349286535", "simultaneous_ci_low": "0.0733096300856565", "simultaneous_ci_high": "0.34894249501829977", "interface": "held-locus common-effect plus cell-contrast readouts", "source_file": "source/onebase/summary.csv", "source_sha256": "9dc5c23c5346f74df63d746469c1258558e043e67b9db877c906c42733252bc6", "selector": "stream=pooled;model=nt100;policy=lambda_1;metric=utility", "source_record": "{\"high95\": \"0.2989373349286535\", \"loci\": \"127\", \"low95\": \"0.12386243944096922\", \"mean\": \"0.21063830067077918\", \"metric\": \"utility\", \"model\": \"nt100\", \"policy\": \"lambda_1\", \"stream\": \"pooled\"}", "note": "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" }, "native_submission": "candidate", "native_value": 0.21063830067077918 }, { "configuration_id": "ntv3_100m_post", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_post", "taskId": "atlas_profile", "value": 0.144, "displayValue": "0.144", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 0.1441415194506628, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_100m_post", "label": "NTv3 100M post", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.1441415194506628", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/existing/summary.csv", "source_sha256": "58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6", "selector": "model=NTv3 100M post;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"NTv3 100M post\", \"repeat1\": \"0.1060014711642299\", \"repeat2\": \"0.1822815677370956\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/existing/summary.csv\", \"source_sha256\": \"58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6\", \"status\": \"independently replayed selection and utility\", \"value\": \"0.1441415194506628\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.1441415194506628 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "prior", "value": 1.492, "displayValue": "1.492", "status": "evaluated", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "sourceValue": 1.492461935201394, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_post/prior", "notes": [ "Same checkpoint-level sequence-prior evaluation as the fitted-head row; no additional model run or fitted prior head.", "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.3927389998597406, "high": 1.5916175825667722, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." }, "manuscriptStatus": "unsupported", "sharedFrom": "ntv3_100m_post", "extensionId": "coverage-20261027" }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "unsupported", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_native;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"\", \"ci_low\": \"\", \"fitting\": \"\", \"higher_is_better\": \"\", \"holdout\": \"\", \"initialization\": \"native\", \"interval\": \"\", \"label\": \"NTv3 100M native head\", \"library\": \"\", \"main_readout\": \"True\", \"main_table_column\": \"prior\", \"metric\": \"\", \"model\": \"ntv3_100m_native\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"\", \"sampling_unit\": \"\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"\", \"source_sha256\": \"\", \"status\": \"unsupported\", \"task\": \"prior\", \"task_label\": \"\", \"value\": \"\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "deduplicated_shared_sequence_prior", "native_value": 1.492461935201394 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_native_outputs_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "opensplice", "value": 0.168, "displayValue": "0.168", "status": "evaluated", "interface": "native annotation proxy", "sourceValue": 0.1677968926004857, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_native/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.1507282603184317, "high": 0.1850513435652891, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.1677968926004857", "ci_low": "0.1507282603184317", "ci_high": "0.1850513435652891", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native annotation proxy", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=ntv3_100m_native", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"ntv3_100m_native\", \"label\": \"NTv3 100M post\\u2020\", \"value\": \"0.1677968926004857\", \"low\": \"0.1507282603184317\", \"high\": \"0.1850513435652891\", \"x\": \"\", \"role\": \"teal\", \"random\": \"False\", \"interface\": \"native annotation proxy\", \"population\": \"590104 records / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"reports/manuscript/figure-data/model-integration-20260917/native/100m/evaluation/opensplice/summary.csv\", \"selector\": \"panel=opensplice/all_interventions;primary=True\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.1677968926004857 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "context", "native_protocol_id": "context_native_outputs_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "context", "value": 0.203, "displayValue": "0.203", "status": "evaluated", "interface": "501-bp DNase log-coverage difference", "sourceValue": 0.202911072445496, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_native/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.1680826996446145, "high": 0.2372679551386709, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.202911072445496", "ci_low": "0.1680826996446145", "ci_high": "0.2372679551386709", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "501-bp DNase log-coverage difference", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_native;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"0.2372679551386709\", \"ci_low\": \"0.1680826996446145\", \"fitting\": \"none; released native head\", \"higher_is_better\": \"True\", \"holdout\": \"fixed original assay population; no new head fitting\", \"initialization\": \"native\", \"interval\": \"95% biological-unit bootstrap\", \"label\": \"NTv3 100M post\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_100m_native\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"Completed September 2026 extension; no held-out labels used for fitting\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"501-bp DNase log-coverage difference\", \"sampling_unit\": \"elements\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"panel=siraj/cell_difference;primary=True\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"reports/manuscript/figure-data/model-integration-20260917/native/100m/evaluation/siraj/summary.csv\", \"source_sha256\": \"56ffd0ba9becdb5796d3f142775d9cff208a90be3175933c2000c12b74f31e01\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context effect contrast\", \"value\": \"0.202911072445496\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.202911072445496 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_native_outputs_accuracy_percent", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "tissue_identification", "value": 27.5, "displayValue": "27.5", "status": "evaluated", "interface": "native RNA outputs", "sourceValue": 27.450980392156865, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_native/tissue_identification", "notes": [ "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "27.450980392156865", "ci_low": "17.647058823529413", "ci_high": "33.33333333333333", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native RNA outputs", "source_file": "source/gtex-results.csv", "source_sha256": "ef1bc015234551618dccc3adb36e0b151ff3f39073667a443c9c7fd03463e921", "selector": "model=ntv3_100m_post;metric=retrieval_top1;scale=100", "source_record": "{\"bootstrap_replicates\": \"10000\", \"high\": \"0.3333333333333333\", \"invalid_replicates\": \"0\", \"low\": \"0.17647058823529413\", \"metric\": \"retrieval_top1\", \"model\": \"ntv3_100m_post\", \"n_genes\": \"5782\", \"n_tissues\": \"51\", \"replicates_with_constant_profiles\": \"0\", \"value\": \"0.27450980392156865\"}", "note": "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" }, "native_submission": "candidate", "native_value": 27.450980392156865 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "background", "native_protocol_id": "background_native_outputs_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "background", "value": 0.152, "displayValue": "0.152", "status": "evaluated", "interface": "501-bp DNase log-coverage difference", "sourceValue": 0.1524855604512107, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_native/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.1230533546930774, "high": 0.181944188703588, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.1524855604512107", "ci_low": "0.1230533546930774", "ci_high": "0.181944188703588", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "501-bp DNase log-coverage difference", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_100m_native;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"0.181944188703588\", \"ci_low\": \"0.1230533546930774\", \"fitting\": \"none; released native head\", \"higher_is_better\": \"True\", \"holdout\": \"fixed original assay population; no new head fitting\", \"initialization\": \"native\", \"interval\": \"95% biological-unit bootstrap\", \"label\": \"NTv3 100M post\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_100m_native\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"Completed September 2026 extension; no held-out labels used for fitting\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"501-bp DNase log-coverage difference\", \"sampling_unit\": \"elements\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"panel=siraj/background_interaction;primary=True\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"reports/manuscript/figure-data/model-integration-20260917/native/100m/evaluation/siraj/summary.csv\", \"source_sha256\": \"56ffd0ba9becdb5796d3f142775d9cff208a90be3175933c2000c12b74f31e01\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj allelic-background interaction\", \"value\": \"0.1524855604512107\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.1524855604512107 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_native_outputs_residual_midrank_percentile", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "repair_percentile", "value": 32.0, "displayValue": "32.0", "status": "evaluated", "interface": "native DNase log coverage, 501bp, frozen primary interface", "sourceValue": 32.012600499400534, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_native/repair_percentile", "notes": [ "331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks" ], "uncertainty": { "level": 0.95, "low": 27.857347018276496, "high": 36.23977632283227, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "32.012600499400534", "ci_low": "27.857347018276496", "ci_high": "36.23977632283227", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native DNase log coverage, 501bp, frozen primary interface", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=ntv3_100m_native;task=repair_percentile", "source_record": "{\"ci_high\": \"36.23977632283227\", \"ci_low\": \"27.857347018276496\", \"interface\": \"native DNase log coverage, 501bp, frozen primary interface\", \"model\": \"ntv3_100m_native\", \"note\": \"331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks\", \"selector\": \"model=ntv3_100m_native\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/repair-audit/table_ready_rows.csv\", \"source_sha256\": \"df33defaa2573bf74488304ac913e68ab338a9a13d818dc7a441648a3c85a44d\", \"task\": \"repair_percentile\", \"value\": \"32.012600499400534\"}", "note": "331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks" }, "native_submission": "candidate", "native_value": 32.012600499400534 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_native_outputs_utility_log2", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "cell_selective", "value": 0.2, "displayValue": "0.200", "status": "evaluated", "interface": "native DNase501 log with held-locus nonnegative per-cell slopes", "sourceValue": 0.2002167361440046, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_native/cell_selective", "notes": [ "Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family" ], "uncertainty": { "level": 0.95, "low": 0.1011433223221313, "high": 0.3017024088346101, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "0.2002167361440046", "ci_low": "0.1011433223221313", "ci_high": "0.3017024088346101", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native DNase501 log with held-locus nonnegative per-cell slopes", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=ntv3_100m_native;task=cell_selective", "source_record": "{\"ci_high\": \"0.3017024088346101\", \"ci_low\": \"0.1011433223221313\", \"interface\": \"native DNase501 log with held-locus nonnegative per-cell slopes\", \"model\": \"ntv3_100m_native\", \"note\": \"Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family\", \"selector\": \"model=ntv3_100m_native\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/ntv3-native/native_table_ready_rows.csv\", \"source_sha256\": \"ae065aa5360faf634d5d582181c9e1d2baa83501143648337b46db2ce3bd0da7\", \"task\": \"cell_selective\", \"value\": \"0.2002167361440046\"}", "note": "Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family" }, "native_submission": "candidate", "native_value": 0.2002167361440046 }, { "configuration_id": "ntv3_100m_native", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_native_outputs_gain_over_random_log2", "checkpoint_target": "InstaDeepAI/NTv3_100M_post", "measurement": { "modelId": "ntv3_100m_native", "taskId": "atlas_profile", "value": 0.344, "displayValue": "0.344", "status": "evaluated", "interface": "source-fitted native functional features", "sourceValue": 0.3436364187186611, "sourceId": "table-1", "sourceRecordId": "ntv3_100m_native/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates.", "Table 1 ‡: a linear head is fitted to native functional outputs; this endpoint is not a direct zero-shot native prediction." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_100m_native", "label": "NTv3 100M post (native)", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.3436364187186611", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted native functional features", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/existing/summary.csv", "source_sha256": "58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6", "selector": "model=NTv3 100M post / Native readout;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted native functional features\", \"low\": \"\", \"model\": \"NTv3 100M post / Native readout\", \"repeat1\": \"0.2923615081644936\", \"repeat2\": \"0.3949113292728287\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/existing/summary.csv\", \"source_sha256\": \"58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6\", \"status\": \"existing score; interface reclassification required\", \"value\": \"0.3436364187186611\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.3436364187186611 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "prior", "value": 1.447, "displayValue": "1.447", "status": "evaluated", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "sourceValue": 1.4470830685902811, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.3524182621796532, "high": 1.5393008696158936, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "evaluated", "value": "1.4470830685902811", "ci_low": "1.3524182621796532", "ci_high": "1.5393008696158936", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_pre;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata_file\\\":\\\"results/prior_benchmark_ntv3_650m_pre_native.json\\\",\\\"panel_sha256\\\":\\\"d4a07c4bcb98c4f5a3f94a6d82a3dfd5e01b3b5baedfc5efa075fe75e3f871f1\\\",\\\"probability_interface\\\":{},\\\"protocol_sha256\\\":\\\"bd8152655f24a60932bd7082b75e6abc3fc855922c1ac15569a2a9e2f1ee2bf2\\\",\\\"source_design\\\":{\\\"architecture\\\":\\\"masked-lm\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"model\\\":\\\"InstaDeepAI/NTv3_650M_pre\\\",\\\"parameter_count\\\":651829435,\\\"random_init\\\":false,\\\"reverse_complement_ensemble\\\":true,\\\"revision\\\":\\\"5e6050bed864a5a8fb32481096bf555495316b31\\\",\\\"score\\\":\\\"mean forward/reverse masked alternate-minus-reference log probability\\\",\\\"seed\\\":20260822,\\\"species\\\":null,\\\"tokenization\\\":\\\"character-level, no special tokens, padded to a multiple of 128\\\"}}\", \"ci_high\": \"1.5393008696158936\", \"ci_low\": \"1.3524182621796532\", \"fitting\": \"Fixed 12% mask rate from the original protocol; no fitted readout or assay labels\", \"higher_is_better\": \"False\", \"holdout\": \"512 fresh 1-Mb blocks; 1,200-bp windows; central base hidden\", \"initialization\": \"trained\", \"interval\": \"95% interval over 512 genomic blocks\", \"label\": \"NTv3 650M pre\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"prior\", \"metric\": \"bits\", \"model\": \"ntv3_650m_pre\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"512\", \"notes\": \"Mean individual-query scores, not a probability ensemble. Same 12% protocol for all NTv3 sizes/stages.\", \"overlay_protocol\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"overlay_selector\": \"figure=prior;panel=A;key=ntv3_650m_pre\", \"overlay_source\": \"reports/manuscript/figure-data/bend2-prior-points.csv\", \"overlay_source_sha256\": \"1937d554c04436e2d3a5888e4414ec66fab230047ee04c164b76aa9b21973ad7\", \"policy\": \"\", \"query\": \"One query per prediction; loss averaged over two fixed mask draws\", \"readout\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"sampling_unit\": \"genomic block\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model_key=ntv3_650m_pre;readout=mask12_single_query\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/prior_calibration.summary.csv\", \"source_sha256\": \"88588953aeeaf592262890b4bc9e6443d855e01a9ba46db91edee5d0e45c42a1\", \"status\": \"evaluated\", \"task\": \"prior\", \"task_label\": \"Hidden-base conditional log loss\", \"value\": \"1.4470830685902811\", \"variant\": \"12% single query\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 1.4470830685902811 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "opensplice", "value": 0.187, "displayValue": "0.187", "status": "evaluated", "interface": "gene-held-out fitted readout", "sourceValue": 0.1869695847355324, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.174298808838127, "high": 0.1994965981503923, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.1869695847355324", "ci_low": "0.174298808838127", "ci_high": "0.1994965981503923", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "gene-held-out fitted readout", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=ntv3_650m_pre", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"ntv3_650m_pre\", \"label\": \"NTv3 650M pre\", \"value\": \"0.1869695847355324\", \"low\": \"0.174298808838127\", \"high\": \"0.1994965981503923\", \"x\": \"\", \"role\": \"blue\", \"random\": \"False\", \"interface\": \"gene-held-out fitted readout\", \"population\": \"590,104 interventions / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"results/opensplice_model_suite.summary.csv\", \"selector\": \"model_key=ntv3_650m_pre;fields=pretrained_*\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.1869695847355324 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "context", "value": 0.01, "displayValue": "0.010", "status": "evaluated", "interface": "Frozen representation + fitted contextual ridge", "sourceValue": 0.010274162383920291, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": -0.0013820938742742962, "high": 0.022521688893341232, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.010274162383920291", "ci_low": "-0.0013820938742742962", "ci_high": "0.022521688893341232", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Frozen representation + fitted contextual ridge", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_pre;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_pre_pretrained.npz\\\",\\\"input_dim\\\":1536,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"masked-lm\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[\\\"InstaDeepAI/ntv3_base_model@0ecff3637f0d3ba5b686d1095083218157c2ca34\\\"],\\\"embedding_dim\\\":1536,\\\"implementation_module\\\":\\\"transformers_modules.InstaDeepAI.ntv3_base_model.0ecff3637f0d3ba5b686d1095083218157c2ca34.modeling_ntv3_pretrained\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/element_split.parquet\\\",\\\"input_sha256\\\":\\\"64941bdb63305fb6555614ef8934f0f563758999c71d01a2ad44e393fe7ae261\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"InstaDeepAI/NTv3_650M_pre\\\",\\\"n_sequences\\\":169912,\\\"pad_to_multiple_of\\\":128,\\\"parameter_count\\\":651829435,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"5e6050bed864a5a8fb32481096bf555495316b31\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"species\\\":null,\\\"state_dict\\\":null,\\\"state_dict_metadata\\\":null},\\\"input_sha256\\\":\\\"5820b9f9d0322f13d90e4acd66f9526232b6df55e925d1734ebb4233de567155\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"01de94b3ea03c8a1b012df50f3d27670d1cdb7d7b6eeb3d25e4b4800cd07efa9\\\",\\\"sign_vector_sha256\\\":\\\"cec65095c95d4e0dbdc85a770336a9cd6e29bb95095b1179ea8cdb179e03ec9f\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":2048,\\\"zero_padding_dimensions\\\":512},\\\"path\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_pre_pretrained_srht512.npz\\\",\\\"sha256\\\":\\\"8842d0ae328894f7f4ba6c70778f3b94e6f75d0c6c9d5990080731da604e44f2\\\"}\", \"ci_high\": \"0.022521688893341232\", \"ci_low\": \"-0.0013820938742742962\", \"fitting\": \"Five-fold element-held-out fitted readout, except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"NTv3 650M pre\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_650m_pre\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Frozen representation + fitted contextual ridge\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=ntv3_650m_pre_pretrained\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_contextual_suite.summary.csv\", \"source_sha256\": \"c948c9ed32febd7e045aa76df7f52cf49da13bed3ecd45e8bb497188a10f3049\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.010274162383920291\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.010274162383920291 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "tissue_identification", "value": 7.8, "displayValue": "7.8", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 7.8431372549019605, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/tissue_identification", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "7.8431372549019605", "ci_low": "3.9215686274509802", "ci_high": "11.76470588235294", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/gtex-nt-results/results.csv", "source_sha256": "7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f", "selector": "model=NTv3 650M pre;endpoint=tissue_id", "source_record": "{\"endpoint\": \"tissue_id\", \"high\": \"11.76470588235294\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"3.9215686274509802\", \"model\": \"NTv3 650M pre\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/gtex-nt-results/results.csv\", \"source_sha256\": \"7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f\", \"status\": \"evaluated; see independent audit\", \"value\": \"7.8431372549019605\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 7.8431372549019605 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "background", "value": 0.009, "displayValue": "0.009", "status": "evaluated", "interface": "Fitted out-of-fold effect differences", "sourceValue": 0.008672778395027, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": -0.0038545004400160694, "high": 0.021358373991477397, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.008672778395027", "ci_low": "-0.0038545004400160694", "ci_high": "0.021358373991477397", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Fitted out-of-fold effect differences", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_pre;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"path\\\":\\\"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_ntv3_650m_pre_pretrained_element_cv.predictions.parquet\\\",\\\"run\\\":{\\\"embedding_metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_pre_pretrained.npz\\\",\\\"input_dim\\\":1536,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"masked-lm\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[\\\"InstaDeepAI/ntv3_base_model@0ecff3637f0d3ba5b686d1095083218157c2ca34\\\"],\\\"embedding_dim\\\":1536,\\\"implementation_module\\\":\\\"transformers_modules.InstaDeepAI.ntv3_base_model.0ecff3637f0d3ba5b686d1095083218157c2ca34.modeling_ntv3_pretrained\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/element_split.parquet\\\",\\\"input_sha256\\\":\\\"64941bdb63305fb6555614ef8934f0f563758999c71d01a2ad44e393fe7ae261\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"InstaDeepAI/NTv3_650M_pre\\\",\\\"n_sequences\\\":169912,\\\"pad_to_multiple_of\\\":128,\\\"parameter_count\\\":651829435,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"5e6050bed864a5a8fb32481096bf555495316b31\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"species\\\":null,\\\"state_dict\\\":null,\\\"state_dict_metadata\\\":null},\\\"input_sha256\\\":\\\"5820b9f9d0322f13d90e4acd66f9526232b6df55e925d1734ebb4233de567155\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"01de94b3ea03c8a1b012df50f3d27670d1cdb7d7b6eeb3d25e4b4800cd07efa9\\\",\\\"sign_vector_sha256\\\":\\\"cec65095c95d4e0dbdc85a770336a9cd6e29bb95095b1179ea8cdb179e03ec9f\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":2048,\\\"zero_padding_dimensions\\\":512},\\\"embedding_path\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_pre_pretrained_srht512.npz\\\",\\\"folds\\\":5,\\\"group_column\\\":\\\"unit_id\\\",\\\"ridge_alpha\\\":10.0,\\\"target\\\":\\\"delta_y\\\"},\\\"sha256\\\":\\\"a26e9b5bbbc1d65e78d439eb0f42c3becfc94624e5ba58c960b975fbd312e71f\\\"}\", \"ci_high\": \"0.021358373991477397\", \"ci_low\": \"-0.0038545004400160694\", \"fitting\": \"Out-of-fold effects with complete elements held out; except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Same mutation on two backgrounds; 114 elements / 68,026 pairs\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"NTv3 650M pre\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_650m_pre\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Fitted out-of-fold effect differences\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=ntv3_650m_pre\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_background_interaction.summary.csv\", \"source_sha256\": \"13fc40cd51049436fb416d7c5c5de74d9c4b5f084be6534276d64ab3e1698c4d\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj sequence-background interaction\", \"value\": \"0.008672778395027\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.008672778395027 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "repair_percentile", "value": 50.4, "displayValue": "50.4", "status": "evaluated", "interface": "held-element response readout", "sourceValue": 50.44972088666388, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/repair_percentile", "notes": [ "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" ], "uncertainty": { "level": 0.95, "low": 47.05829930681474, "high": 53.86714612200989, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "50.44972088666388", "ci_low": "47.05829930681474", "ci_high": "53.86714612200989", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "held-element response readout", "source_file": "source/repair-summary.csv", "source_sha256": "433befcbc4b8c0d5d7b3779c2a2aa1c15257bd5a4cd28aaeb35ae0ccbd1ca5a7", "selector": "model=ntv3_650m_pre;value=100*candidate_percentile", "source_record": "{\"candidate_percentile\": \"0.5044972088666388\", \"ci_high\": \"0.5386714612200989\", \"ci_low\": \"0.4705829930681474\", \"elements\": \"110\", \"family\": \"foundation\", \"label\": \"NTv3 650M pre\", \"model\": \"ntv3_650m_pre\", \"resolved_instances\": \"331\", \"top1_success_ci_high\": \"0.11136363636363636\", \"top1_success_ci_low\": \"0.04545454545454545\", \"top1_success_rate\": \"0.07651515151515151\"}", "note": "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" }, "native_submission": "candidate", "native_value": 50.44972088666388 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "cell_selective", "value": -0.061, "displayValue": "-0.061", "status": "evaluated", "interface": "fitted common+contrast, whole-locus OOF", "sourceValue": -0.0614350486039921, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/cell_selective", "notes": [ "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." ], "uncertainty": { "level": 0.95, "low": -0.0944439042024492, "high": -0.0340523391328295, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "-0.0614350486039921", "ci_low": "-0.0944439042024492", "ci_high": "-0.0340523391328295", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted common+contrast, whole-locus OOF", "source_file": "source/missing-model-evaluations-20260922.csv", "source_sha256": "44aba49df3213d71cee00ae657e14769228944b01dfee48e6ef487cbcea0a818", "selector": "model=ntv3_650m_pre;task=cell_selective", "source_record": "{\"ci_high\": \"-0.0340523391328295\", \"ci_low\": \"-0.0944439042024492\", \"interface\": \"fitted common+contrast, whole-locus OOF\", \"interval\": \"pointwise 95% whole-locus bootstrap, fixed predictions\", \"model\": \"ntv3_650m_pre\", \"n_groups\": \"127\", \"n_units\": \"280\", \"result_file\": \"output/missing-cellselective-evaluations-20260922/cached-four/table_ready_rows.csv\", \"result_sha256\": \"dd125a82b3ff26b8c9bcc9318aeb8562f84226b3248993627fc369a6850fc58c\", \"source_path\": \"/Users/fvk220/Documents/New project/output/missing-cellselective-evaluations-20260922/cached-four/summary.csv\", \"source_sha256\": \"220905a9d94b369a0350a642248012db11264ea09ccd88edc122c34a9be4b902\", \"task\": \"cell_selective\", \"value\": \"-0.0614350486039921\"}", "note": "Audited retrospective coverage extension; unchanged cohorts, held-locus folds and task definitions. Added intervals are pointwise, conditional on fixed fits and nominees; added design models are outside the original 24-contrast family. Readout widths documented in Appendix cell-selective methods." }, "native_submission": "candidate", "native_value": -0.0614350486039921 }, { "configuration_id": "ntv3_650m_pre", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "InstaDeepAI/NTv3_650M_pre", "measurement": { "modelId": "ntv3_650m_pre", "taskId": "atlas_profile", "value": 0.368, "displayValue": "0.368", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 0.3679799614366345, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_pre/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_650m_pre", "label": "NTv3 650M pre", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.3679799614366345", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/five-models/summary.csv", "source_sha256": "e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092", "selector": "model=NTv3 650M pre;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"NTv3 650M pre\", \"repeat1\": \"0.390837783518738\", \"repeat2\": \"0.3451221393545309\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/five-models/summary.csv\", \"source_sha256\": \"e3ac4f031ff4555dd8fe9ca95c29b8a5e6eca99d8946bdc9af3d39db24aed092\", \"status\": \"independently replayed selection and utility\", \"value\": \"0.3679799614366345\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.3679799614366345 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "prior", "value": 1.375, "displayValue": "1.375", "status": "evaluated", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "sourceValue": 1.3745521953833992, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.274789743100344, "high": 1.47500625212311, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "evaluated", "value": "1.3745521953833992", "ci_low": "1.2747897431003441", "ci_high": "1.47500625212311", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_post;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata_file\\\":\\\"results/prior_benchmark_ntv3_650m_post_native.json\\\",\\\"panel_sha256\\\":\\\"d4a07c4bcb98c4f5a3f94a6d82a3dfd5e01b3b5baedfc5efa075fe75e3f871f1\\\",\\\"probability_interface\\\":{},\\\"protocol_sha256\\\":\\\"bd8152655f24a60932bd7082b75e6abc3fc855922c1ac15569a2a9e2f1ee2bf2\\\",\\\"source_design\\\":{\\\"architecture\\\":\\\"base\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"model\\\":\\\"InstaDeepAI/NTv3_650M_post\\\",\\\"parameter_count\\\":679787252,\\\"random_init\\\":false,\\\"reverse_complement_ensemble\\\":true,\\\"revision\\\":\\\"ad622051abbe9376bdea3f2727863fa559957e9f\\\",\\\"score\\\":\\\"mean forward/reverse masked alternate-minus-reference log probability\\\",\\\"seed\\\":20260822,\\\"species\\\":\\\"human\\\",\\\"tokenization\\\":\\\"character-level, no special tokens, padded to a multiple of 128\\\"}}\", \"ci_high\": \"1.47500625212311\", \"ci_low\": \"1.2747897431003441\", \"fitting\": \"Fixed 12% mask rate from the original protocol; no fitted readout or assay labels\", \"higher_is_better\": \"False\", \"holdout\": \"512 fresh 1-Mb blocks; 1,200-bp windows; central base hidden\", \"initialization\": \"trained\", \"interval\": \"95% interval over 512 genomic blocks\", \"label\": \"NTv3 650M post\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"prior\", \"metric\": \"bits\", \"model\": \"ntv3_650m_post\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"512\", \"notes\": \"Mean individual-query scores, not a probability ensemble. Same 12% protocol for all NTv3 sizes/stages.\", \"overlay_protocol\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"overlay_selector\": \"figure=prior;panel=A;key=ntv3_650m_post\", \"overlay_source\": \"reports/manuscript/figure-data/bend2-prior-points.csv\", \"overlay_source_sha256\": \"1937d554c04436e2d3a5888e4414ec66fab230047ee04c164b76aa9b21973ad7\", \"policy\": \"\", \"query\": \"One query per prediction; loss averaged over two fixed mask draws\", \"readout\": \"fixed 12% mask; mean individual-query loss across two draws; no ensemble\", \"sampling_unit\": \"genomic block\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model_key=ntv3_650m_post;readout=mask12_single_query\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/prior_calibration.summary.csv\", \"source_sha256\": \"88588953aeeaf592262890b4bc9e6443d855e01a9ba46db91edee5d0e45c42a1\", \"status\": \"evaluated\", \"task\": \"prior\", \"task_label\": \"Hidden-base conditional log loss\", \"value\": \"1.3745521953833992\", \"variant\": \"12% single query\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 1.3745521953833992 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "opensplice", "value": 0.564, "displayValue": "0.564", "status": "evaluated", "interface": "gene-held-out fitted readout", "sourceValue": 0.5643109042285411, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.5499994852502554, "high": 0.5784303144690341, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.5643109042285411", "ci_low": "0.5499994852502554", "ci_high": "0.5784303144690341", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "gene-held-out fitted readout", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=ntv3_650m_post", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"ntv3_650m_post\", \"label\": \"NTv3 650M post\", \"value\": \"0.5643109042285411\", \"low\": \"0.5499994852502554\", \"high\": \"0.5784303144690341\", \"x\": \"\", \"role\": \"teal\", \"random\": \"False\", \"interface\": \"gene-held-out fitted readout\", \"population\": \"590,104 interventions / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"results/opensplice_model_suite.summary.csv\", \"selector\": \"model_key=ntv3_650m_post;fields=pretrained_*\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.5643109042285411 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "context", "value": 0.256, "displayValue": "0.256", "status": "evaluated", "interface": "Frozen representation + fitted contextual ridge", "sourceValue": 0.2559890462602803, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.22877624701877824, "high": 0.2833694522417906, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.2559890462602803", "ci_low": "0.22877624701877824", "ci_high": "0.2833694522417906", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Frozen representation + fitted contextual ridge", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_post;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_post_pretrained.npz\\\",\\\"input_dim\\\":1536,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"base\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[\\\"InstaDeepAI/ntv3_base_model@0ecff3637f0d3ba5b686d1095083218157c2ca34\\\"],\\\"embedding_dim\\\":1536,\\\"implementation_module\\\":\\\"transformers_modules.InstaDeepAI.ntv3_base_model.0ecff3637f0d3ba5b686d1095083218157c2ca34.modeling_ntv3_posttrained\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/element_split.parquet\\\",\\\"input_sha256\\\":\\\"64941bdb63305fb6555614ef8934f0f563758999c71d01a2ad44e393fe7ae261\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"InstaDeepAI/NTv3_650M_post\\\",\\\"n_sequences\\\":169912,\\\"pad_to_multiple_of\\\":128,\\\"parameter_count\\\":679787252,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"ad622051abbe9376bdea3f2727863fa559957e9f\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"species\\\":\\\"human\\\",\\\"state_dict\\\":null,\\\"state_dict_metadata\\\":null},\\\"input_sha256\\\":\\\"21edcec94b368b64fc1e793ed09d212e128eb876846e674291b727d7c6564868\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"01de94b3ea03c8a1b012df50f3d27670d1cdb7d7b6eeb3d25e4b4800cd07efa9\\\",\\\"sign_vector_sha256\\\":\\\"cec65095c95d4e0dbdc85a770336a9cd6e29bb95095b1179ea8cdb179e03ec9f\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":2048,\\\"zero_padding_dimensions\\\":512},\\\"path\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_post_pretrained_srht512.npz\\\",\\\"sha256\\\":\\\"932d863c86ef9c9e702bc834d50881676928761709fb7e68c49175b4947b848a\\\"}\", \"ci_high\": \"0.2833694522417906\", \"ci_low\": \"0.22877624701877824\", \"fitting\": \"Five-fold element-held-out fitted readout, except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"NTv3 650M post\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_650m_post\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Frozen representation + fitted contextual ridge\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=ntv3_650m_post_pretrained\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_contextual_suite.summary.csv\", \"source_sha256\": \"c948c9ed32febd7e045aa76df7f52cf49da13bed3ecd45e8bb497188a10f3049\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.2559890462602803\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.2559890462602803 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "tissue_identification", "value": 56.9, "displayValue": "56.9", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 56.86274509803921, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/tissue_identification", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "56.86274509803921", "ci_low": "45.09803921568627", "ci_high": "62.745098039215684", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/gtex-nt-results/results.csv", "source_sha256": "7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f", "selector": "model=NTv3 650M post;endpoint=tissue_id", "source_record": "{\"endpoint\": \"tissue_id\", \"high\": \"62.745098039215684\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"45.09803921568627\", \"model\": \"NTv3 650M post\", \"repeat1\": \"\", \"repeat2\": \"\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/gtex-nt-results/results.csv\", \"source_sha256\": \"7958306edd5e637e1163df844c8764fc8162ffe3ef082fd30232fd87c5a3fa7f\", \"status\": \"evaluated; see independent audit\", \"value\": \"56.86274509803921\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family." }, "native_submission": "candidate", "native_value": 56.86274509803921 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "background", "value": 0.187, "displayValue": "0.187", "status": "evaluated", "interface": "Fitted out-of-fold effect differences", "sourceValue": 0.18701246956946582, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.15459263193225128, "high": 0.21975255139994365, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.18701246956946582", "ci_low": "0.15459263193225128", "ci_high": "0.21975255139994365", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Fitted out-of-fold effect differences", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_post;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"path\\\":\\\"/e/project1/e-dev-2026d02-064/_abj/gimbench/results/siraj_ntv3_650m_post_pretrained_element_cv.predictions.parquet\\\",\\\"run\\\":{\\\"embedding_metadata\\\":{\\\"fit_to_benchmark_sequences\\\":false,\\\"input\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_post_pretrained.npz\\\",\\\"input_dim\\\":1536,\\\"input_metadata\\\":{\\\"architecture\\\":\\\"base\\\",\\\"code_revision\\\":\\\"0ecff3637f0d3ba5b686d1095083218157c2ca34\\\",\\\"custom_flash_attention_disabled\\\":false,\\\"dependencies\\\":[\\\"InstaDeepAI/ntv3_base_model@0ecff3637f0d3ba5b686d1095083218157c2ca34\\\"],\\\"embedding_dim\\\":1536,\\\"implementation_module\\\":\\\"transformers_modules.InstaDeepAI.ntv3_base_model.0ecff3637f0d3ba5b686d1095083218157c2ca34.modeling_ntv3_posttrained\\\",\\\"input\\\":\\\"data/processed/siraj_satmut/element_split.parquet\\\",\\\"input_sha256\\\":\\\"64941bdb63305fb6555614ef8934f0f563758999c71d01a2ad44e393fe7ae261\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"InstaDeepAI/NTv3_650M_post\\\",\\\"n_sequences\\\":169912,\\\"pad_to_multiple_of\\\":128,\\\"parameter_count\\\":679787252,\\\"pooling\\\":\\\"mean_non_special_tokens\\\",\\\"random_init\\\":false,\\\"revision\\\":\\\"ad622051abbe9376bdea3f2727863fa559957e9f\\\",\\\"seed\\\":20260822,\\\"sequence_columns\\\":[\\\"sequence_ref\\\",\\\"sequence_alt\\\"],\\\"species\\\":\\\"human\\\",\\\"state_dict\\\":null,\\\"state_dict_metadata\\\":null},\\\"input_sha256\\\":\\\"21edcec94b368b64fc1e793ed09d212e128eb876846e674291b727d7c6564868\\\",\\\"label_independent\\\":true,\\\"n_sequences\\\":169912,\\\"output_dim\\\":512,\\\"projection\\\":\\\"seeded sign flip, unnormalized Walsh-Hadamard transform, and fixed coordinate subsampling\\\",\\\"projection_scaling\\\":\\\"1/sqrt(output_dim)\\\",\\\"seed\\\":20260822,\\\"selected_coordinate_sha256\\\":\\\"01de94b3ea03c8a1b012df50f3d27670d1cdb7d7b6eeb3d25e4b4800cd07efa9\\\",\\\"sign_vector_sha256\\\":\\\"cec65095c95d4e0dbdc85a770336a9cd6e29bb95095b1179ea8cdb179e03ec9f\\\",\\\"storage_dtype\\\":\\\"float16\\\",\\\"transform_dim\\\":2048,\\\"zero_padding_dimensions\\\":512},\\\"embedding_path\\\":\\\"data/cache/embeddings/siraj_ntv3_650m_post_pretrained_srht512.npz\\\",\\\"folds\\\":5,\\\"group_column\\\":\\\"unit_id\\\",\\\"ridge_alpha\\\":10.0,\\\"target\\\":\\\"delta_y\\\"},\\\"sha256\\\":\\\"facacaa984cfb85de1d76fc44dae6d68ff2ceaea925c7bc2e6ae2e47623a7cf8\\\"}\", \"ci_high\": \"0.21975255139994365\", \"ci_low\": \"0.15459263193225128\", \"fitting\": \"Out-of-fold effects with complete elements held out; except native specialists\", \"higher_is_better\": \"True\", \"holdout\": \"Same mutation on two backgrounds; 114 elements / 68,026 pairs\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"NTv3 650M post\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_650m_post\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Fitted out-of-fold effect differences\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=ntv3_650m_post\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_background_interaction.summary.csv\", \"source_sha256\": \"13fc40cd51049436fb416d7c5c5de74d9c4b5f084be6534276d64ab3e1698c4d\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj sequence-background interaction\", \"value\": \"0.18701246956946582\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.18701246956946582 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "repair_percentile", "value": 30.0, "displayValue": "30.0", "status": "evaluated", "interface": "held-element response readout", "sourceValue": 30.04843645057621, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/repair_percentile", "notes": [ "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" ], "uncertainty": { "level": 0.95, "low": 25.76013630232495, "high": 34.549587462636005, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "30.04843645057621", "ci_low": "25.76013630232495", "ci_high": "34.549587462636005", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "held-element response readout", "source_file": "source/repair-summary.csv", "source_sha256": "433befcbc4b8c0d5d7b3779c2a2aa1c15257bd5a4cd28aaeb35ae0ccbd1ca5a7", "selector": "model=ntv3_650m_post;value=100*candidate_percentile", "source_record": "{\"candidate_percentile\": \"0.3004843645057621\", \"ci_high\": \"0.34549587462636006\", \"ci_low\": \"0.2576013630232495\", \"elements\": \"110\", \"family\": \"functional\", \"label\": \"NTv3 650M post\", \"model\": \"ntv3_650m_post\", \"resolved_instances\": \"331\", \"top1_success_ci_high\": \"0.26439393939393935\", \"top1_success_ci_low\": \"0.16287878787878785\", \"top1_success_rate\": \"0.2121212121212121\"}", "note": "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" }, "native_submission": "candidate", "native_value": 30.04843645057621 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "cell_selective", "value": 0.287, "displayValue": "0.287", "status": "evaluated", "interface": "held-locus common-effect plus cell-contrast readouts", "sourceValue": 0.28730864470401507, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/cell_selective", "notes": [ "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" ], "uncertainty": { "level": 0.95, "low": 0.19782432693585955, "high": 0.37734504266748736, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "0.28730864470401507", "ci_low": "0.19782432693585955", "ci_high": "0.37734504266748736", "simultaneous_ci_low": "0.14434480887008558", "simultaneous_ci_high": "0.4288643419279185", "interface": "held-locus common-effect plus cell-contrast readouts", "source_file": "source/onebase/summary.csv", "source_sha256": "9dc5c23c5346f74df63d746469c1258558e043e67b9db877c906c42733252bc6", "selector": "stream=pooled;model=nt650;policy=lambda_1;metric=utility", "source_record": "{\"high95\": \"0.37734504266748736\", \"loci\": \"127\", \"low95\": \"0.19782432693585955\", \"mean\": \"0.28730864470401507\", \"metric\": \"utility\", \"model\": \"nt650\", \"policy\": \"lambda_1\", \"stream\": \"pooled\"}", "note": "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" }, "native_submission": "candidate", "native_value": 0.28730864470401507 }, { "configuration_id": "ntv3_650m_post", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_post", "taskId": "atlas_profile", "value": 0.568, "displayValue": "0.568", "status": "evaluated", "interface": "source-fitted frozen hidden representation", "sourceValue": 0.5684901276910762, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_650m_post", "label": "NTv3 650M post", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.5684901276910762", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted frozen hidden representation", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/existing/summary.csv", "source_sha256": "58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6", "selector": "model=NTv3 650M post;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted frozen hidden representation\", \"low\": \"\", \"model\": \"NTv3 650M post\", \"repeat1\": \"0.5267446738214636\", \"repeat2\": \"0.6102355815606888\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/existing/summary.csv\", \"source_sha256\": \"58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6\", \"status\": \"independently replayed selection and utility\", \"value\": \"0.5684901276910762\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.5684901276910762 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "prior", "value": 1.375, "displayValue": "1.375", "status": "evaluated", "interface": "fixed 12% mask; mean individual-query loss across two draws; no ensemble", "sourceValue": 1.3745521953833992, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_post/prior", "notes": [ "Same checkpoint-level sequence-prior evaluation as the fitted-head row; no additional model run or fitted prior head.", "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 1.274789743100344, "high": 1.47500625212311, "type": "reported pointwise bootstrap interval", "samplingUnit": "genomic block", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." }, "manuscriptStatus": "unsupported", "sharedFrom": "ntv3_650m_post", "extensionId": "coverage-20261027" }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "unsupported", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_native;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"\", \"ci_low\": \"\", \"fitting\": \"\", \"higher_is_better\": \"\", \"holdout\": \"\", \"initialization\": \"native\", \"interval\": \"\", \"label\": \"NTv3 650M native head\", \"library\": \"\", \"main_readout\": \"True\", \"main_table_column\": \"prior\", \"metric\": \"\", \"model\": \"ntv3_650m_native\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"\", \"sampling_unit\": \"\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"\", \"source_sha256\": \"\", \"status\": \"unsupported\", \"task\": \"prior\", \"task_label\": \"\", \"value\": \"\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "deduplicated_shared_sequence_prior", "native_value": 1.3745521953833992 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_native_outputs_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "opensplice", "value": 0.461, "displayValue": "0.461", "status": "evaluated", "interface": "native annotation proxy", "sourceValue": 0.4613678400264601, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_native/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.4453194331063385, "high": 0.4773536878840134, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.4613678400264601", "ci_low": "0.4453194331063385", "ci_high": "0.4773536878840134", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native annotation proxy", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=ntv3_650m_native", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"ntv3_650m_native\", \"label\": \"NTv3 650M post\\u2020\", \"value\": \"0.4613678400264601\", \"low\": \"0.4453194331063385\", \"high\": \"0.4773536878840134\", \"x\": \"\", \"role\": \"teal\", \"random\": \"False\", \"interface\": \"native annotation proxy\", \"population\": \"590104 records / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"reports/manuscript/figure-data/model-integration-20260917/native/650m/evaluation/opensplice/summary.csv\", \"selector\": \"panel=opensplice/all_interventions;primary=True\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.4613678400264601 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "context", "native_protocol_id": "context_native_outputs_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "context", "value": 0.239, "displayValue": "0.239", "status": "evaluated", "interface": "501-bp DNase log-coverage difference", "sourceValue": 0.2394553535716708, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_native/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.2073213297522291, "high": 0.2711241643711061, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.2394553535716708", "ci_low": "0.2073213297522291", "ci_high": "0.2711241643711061", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "501-bp DNase log-coverage difference", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_native;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"0.2711241643711061\", \"ci_low\": \"0.2073213297522291\", \"fitting\": \"none; released native head\", \"higher_is_better\": \"True\", \"holdout\": \"fixed original assay population; no new head fitting\", \"initialization\": \"native\", \"interval\": \"95% biological-unit bootstrap\", \"label\": \"NTv3 650M post\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_650m_native\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"Completed September 2026 extension; no held-out labels used for fitting\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"501-bp DNase log-coverage difference\", \"sampling_unit\": \"elements\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"panel=siraj/cell_difference;primary=True\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"reports/manuscript/figure-data/model-integration-20260917/native/650m/evaluation/siraj/summary.csv\", \"source_sha256\": \"4e52dcad8da2d14139b4850b9e41ce8bb78d81d06a15a472de74fd60a991d8f6\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context effect contrast\", \"value\": \"0.2394553535716708\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.2394553535716708 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_native_outputs_accuracy_percent", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "tissue_identification", "value": 45.1, "displayValue": "45.1", "status": "evaluated", "interface": "native RNA outputs", "sourceValue": 45.09803921568628, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_native/tissue_identification", "notes": [ "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "45.09803921568628", "ci_low": "35.294117647058826", "ci_high": "47.05882352941176", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native RNA outputs", "source_file": "source/gtex-results.csv", "source_sha256": "ef1bc015234551618dccc3adb36e0b151ff3f39073667a443c9c7fd03463e921", "selector": "model=ntv3_650m_post;metric=retrieval_top1;scale=100", "source_record": "{\"bootstrap_replicates\": \"10000\", \"high\": \"0.47058823529411764\", \"invalid_replicates\": \"0\", \"low\": \"0.35294117647058826\", \"metric\": \"retrieval_top1\", \"model\": \"ntv3_650m_post\", \"n_genes\": \"5782\", \"n_tissues\": \"51\", \"replicates_with_constant_profiles\": \"0\", \"value\": \"0.45098039215686275\"}", "note": "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" }, "native_submission": "candidate", "native_value": 45.09803921568628 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "background", "native_protocol_id": "background_native_outputs_spearman_rho", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "background", "value": 0.17, "displayValue": "0.170", "status": "evaluated", "interface": "501-bp DNase log-coverage difference", "sourceValue": 0.1698682962800696, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_native/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.138192262168277, "high": 0.200875785992956, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.1698682962800696", "ci_low": "0.138192262168277", "ci_high": "0.200875785992956", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "501-bp DNase log-coverage difference", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=ntv3_650m_native;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"0.200875785992956\", \"ci_low\": \"0.138192262168277\", \"fitting\": \"none; released native head\", \"higher_is_better\": \"True\", \"holdout\": \"fixed original assay population; no new head fitting\", \"initialization\": \"native\", \"interval\": \"95% biological-unit bootstrap\", \"label\": \"NTv3 650M post\", \"library\": \"\", \"main_readout\": \"False\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"ntv3_650m_native\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"Completed September 2026 extension; no held-out labels used for fitting\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"501-bp DNase log-coverage difference\", \"sampling_unit\": \"elements\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"panel=siraj/background_interaction;primary=True\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"reports/manuscript/figure-data/model-integration-20260917/native/650m/evaluation/siraj/summary.csv\", \"source_sha256\": \"4e52dcad8da2d14139b4850b9e41ce8bb78d81d06a15a472de74fd60a991d8f6\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj allelic-background interaction\", \"value\": \"0.1698682962800696\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.1698682962800696 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_native_outputs_residual_midrank_percentile", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "repair_percentile", "value": 29.3, "displayValue": "29.3", "status": "evaluated", "interface": "native DNase log coverage, 501bp, frozen primary interface", "sourceValue": 29.31304047963624, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_native/repair_percentile", "notes": [ "331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks" ], "uncertainty": { "level": 0.95, "low": 25.52074138048832, "high": 33.272651059675, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "29.31304047963624", "ci_low": "25.52074138048832", "ci_high": "33.272651059675", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native DNase log coverage, 501bp, frozen primary interface", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=ntv3_650m_native;task=repair_percentile", "source_record": "{\"ci_high\": \"33.272651059675\", \"ci_low\": \"25.52074138048832\", \"interface\": \"native DNase log coverage, 501bp, frozen primary interface\", \"model\": \"ntv3_650m_native\", \"note\": \"331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks\", \"selector\": \"model=ntv3_650m_native\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/repair-audit/table_ready_rows.csv\", \"source_sha256\": \"df33defaa2573bf74488304ac913e68ab338a9a13d818dc7a441648a3c85a44d\", \"task\": \"repair_percentile\", \"value\": \"29.31304047963624\"}", "note": "331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks" }, "native_submission": "candidate", "native_value": 29.31304047963624 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_native_outputs_utility_log2", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "cell_selective", "value": 0.25, "displayValue": "0.250", "status": "evaluated", "interface": "native DNase501 log with held-locus nonnegative per-cell slopes", "sourceValue": 0.2503545904194044, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_native/cell_selective", "notes": [ "Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family" ], "uncertainty": { "level": 0.95, "low": 0.1623620369881065, "high": 0.341042496773598, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "0.2503545904194044", "ci_low": "0.1623620369881065", "ci_high": "0.341042496773598", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native DNase501 log with held-locus nonnegative per-cell slopes", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=ntv3_650m_native;task=cell_selective", "source_record": "{\"ci_high\": \"0.341042496773598\", \"ci_low\": \"0.1623620369881065\", \"interface\": \"native DNase501 log with held-locus nonnegative per-cell slopes\", \"model\": \"ntv3_650m_native\", \"note\": \"Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family\", \"selector\": \"model=ntv3_650m_native\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/ntv3-native/native_table_ready_rows.csv\", \"source_sha256\": \"ae065aa5360faf634d5d582181c9e1d2baa83501143648337b46db2ce3bd0da7\", \"task\": \"cell_selective\", \"value\": \"0.2503545904194044\"}", "note": "Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family" }, "native_submission": "candidate", "native_value": 0.2503545904194044 }, { "configuration_id": "ntv3_650m_native", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_native_outputs_gain_over_random_log2", "checkpoint_target": "InstaDeepAI/NTv3_650M_post", "measurement": { "modelId": "ntv3_650m_native", "taskId": "atlas_profile", "value": 0.482, "displayValue": "0.482", "status": "evaluated", "interface": "source-fitted native functional features", "sourceValue": 0.4820916958518555, "sourceId": "table-1", "sourceRecordId": "ntv3_650m_native/atlas_profile", "notes": [ "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates.", "Table 1 ‡: a linear head is fitted to native functional outputs; this endpoint is not a direct zero-shot native prediction." ], "uncertainty": null }, "archived_source_record": { "model": "ntv3_650m_native", "label": "NTv3 650M post (native)", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.4820916958518555", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-fitted native functional features", "source_file": "provenance/table-completion-20260923/sources/table-completion-20260923/atlas/evaluations/existing/summary.csv", "source_sha256": "58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6", "selector": "model=NTv3 650M post / Native readout;endpoint=multi_cell_gain", "source_record": "{\"endpoint\": \"multi_cell_gain\", \"high\": \"\", \"interface\": \"source-fitted native functional features\", \"low\": \"\", \"model\": \"NTv3 650M post / Native readout\", \"repeat1\": \"0.4302441032390426\", \"repeat2\": \"0.5339392884646684\", \"source\": \"/Users/fvk220/Documents/New project/output/table-completion-20260923/atlas/evaluations/existing/summary.csv\", \"source_sha256\": \"58a326e626c64f6101e6b477cbc3cf0dcb1bed7b5fc2b47b27159b8b9b531bf6\", \"status\": \"existing score; interface reclassification required\", \"value\": \"0.4820916958518555\"}", "note": "Retrospective model-roster extension of frozen task; source-only fitting. Pointwise intervals; no expanded confirmatory family. Mean of two RNA-repeat gains over the same 16 requests; repeats are descriptive, not independent uncertainty estimates." }, "native_submission": "candidate", "native_value": 0.4820916958518555 }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "prior", "value": null, "displayValue": "—", "status": "unsupported", "interface": null, "sourceValue": null, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/prior", "notes": [], "uncertainty": null, "manuscriptStatus": "unevaluated", "missingReason": "AlphaGenome does not provide nucleotide probabilities for this hidden-base task. Fitting its functional readouts does not supply a sequence-prior head.", "extensionId": "coverage-20261027" }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "unevaluated", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "", "source_sha256": "", "selector": "", "source_record": "", "note": "" }, "native_submission": "omitted_missing_or_unsupported_never_zero" }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_embeddings_spearman_rho", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "opensplice", "value": 0.65, "displayValue": "0.650", "status": "evaluated", "interface": "frozen hidden-representation fitted linear readout", "sourceValue": 0.6498792466499858, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A.", "Interface resolved from the Table 1 caption because the CSV interface field is blank." ], "uncertainty": { "level": 0.95, "low": 0.6343857827907978, "high": 0.6650260674824582, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.6498792466499858", "ci_low": "0.6343857827907978", "ci_high": "0.6650260674824582", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=alphagenome_fitted", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"alphagenome_fitted\", \"label\": \"AlphaGenome \\u00b7 fitted\", \"value\": \"0.6498792466499858\", \"low\": \"0.6343857827907978\", \"high\": \"0.6650260674824582\", \"x\": \"\", \"role\": \"teal\", \"random\": \"False\", \"interface\": \"\", \"population\": \"\", \"interval\": \"\", \"source\": \"\", \"selector\": \"\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.6498792466499858 }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "context", "native_protocol_id": "context_embeddings_spearman_rho", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "context", "value": 0.402, "displayValue": "0.402", "status": "evaluated", "interface": "matched assay insert hidden-representation ridge readout", "sourceValue": 0.4021855551043893, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/context", "notes": [ "Matched held-out identities, truth and biological units audited against NTv3; 10000 whole-unit bootstrap replicates; strict transfer has 88 target genes, 86 correlation-eligible genes for all models" ], "uncertainty": { "level": 0.95, "low": 0.37140027981142143, "high": 0.43348611534948495, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.4021855551043893", "ci_low": "0.37140027981142143", "ci_high": "0.43348611534948495", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "matched assay insert hidden-representation ridge readout", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=alphagenome_fitted;task=context", "source_record": "{\"ci_high\": \"0.43348611534948495\", \"ci_low\": \"0.37140027981142143\", \"interface\": \"matched assay insert hidden-representation ridge readout\", \"model\": \"alphagenome_fitted\", \"note\": \"Matched held-out identities, truth and biological units audited against NTv3; 10000 whole-unit bootstrap replicates; strict transfer has 88 target genes, 86 correlation-eligible genes for all models\", \"selector\": \"task=siraj_cell_contrast;model=alphagenome_fitted\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/matched-audit/summary.csv\", \"source_sha256\": \"544fe2105a58f59d79fbf1e53292356dd4f2339734f16dc09687a2e338b0d8b6\", \"task\": \"context\", \"value\": \"0.4021855551043893\"}", "note": "Matched held-out identities, truth and biological units audited against NTv3; 10000 whole-unit bootstrap replicates; strict transfer has 88 target genes, 86 correlation-eligible genes for all models" }, "native_submission": "candidate", "native_value": 0.4021855551043893 }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_embeddings_accuracy_percent", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "tissue_identification", "value": 64.7, "displayValue": "64.7", "status": "evaluated", "interface": "GTEx-fitted genomic hidden-representation ridge readout", "sourceValue": 64.70588235294117, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/tissue_identification", "notes": [ "5782 common held-out genes and 51 tissues; independent rank replay and bootstrap validation passed" ], "uncertainty": null }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "64.70588235294117", "ci_low": "50.98039215686274", "ci_high": "68.62745098039215", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "GTEx-fitted genomic hidden-representation ridge readout", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=alphagenome_fitted;task=tissue_identification", "source_record": "{\"ci_high\": \"68.62745098039215\", \"ci_low\": \"50.98039215686274\", \"interface\": \"GTEx-fitted genomic hidden-representation ridge readout\", \"model\": \"alphagenome_fitted\", \"note\": \"5782 common held-out genes and 51 tissues; independent rank replay and bootstrap validation passed\", \"selector\": \"metric=retrieval_top1;scale=100\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/recovered/gtex-fitted-results.csv\", \"source_sha256\": \"d71e51d1e809abbddfee2956ff08246f8cc594270dc3b2c9d67a821828bac25e\", \"task\": \"tissue_identification\", \"value\": \"64.70588235294117\"}", "note": "5782 common held-out genes and 51 tissues; independent rank replay and bootstrap validation passed" }, "native_submission": "candidate", "native_value": 64.70588235294117 }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "background", "native_protocol_id": "background_embeddings_spearman_rho", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "background", "value": 0.269, "displayValue": "0.269", "status": "evaluated", "interface": "matched assay insert hidden-representation ridge readout", "sourceValue": 0.26862234598973245, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/background", "notes": [ "Matched held-out identities, truth and biological units audited against NTv3; 10000 whole-unit bootstrap replicates; strict transfer has 88 target genes, 86 correlation-eligible genes for all models" ], "uncertainty": { "level": 0.95, "low": 0.23246584867660344, "high": 0.30428710025153033, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.26862234598973245", "ci_low": "0.23246584867660344", "ci_high": "0.30428710025153033", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "matched assay insert hidden-representation ridge readout", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=alphagenome_fitted;task=background", "source_record": "{\"ci_high\": \"0.30428710025153033\", \"ci_low\": \"0.23246584867660344\", \"interface\": \"matched assay insert hidden-representation ridge readout\", \"model\": \"alphagenome_fitted\", \"note\": \"Matched held-out identities, truth and biological units audited against NTv3; 10000 whole-unit bootstrap replicates; strict transfer has 88 target genes, 86 correlation-eligible genes for all models\", \"selector\": \"task=siraj_background;model=alphagenome_fitted\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/matched-audit/summary.csv\", \"source_sha256\": \"544fe2105a58f59d79fbf1e53292356dd4f2339734f16dc09687a2e338b0d8b6\", \"task\": \"background\", \"value\": \"0.26862234598973245\"}", "note": "Matched held-out identities, truth and biological units audited against NTv3; 10000 whole-unit bootstrap replicates; strict transfer has 88 target genes, 86 correlation-eligible genes for all models" }, "native_submission": "candidate", "native_value": 0.26862234598973245 }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_embeddings_residual_midrank_percentile", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "repair_percentile", "value": 25.7, "displayValue": "25.7", "status": "evaluated", "interface": "matched insert hidden-feature OOF ridge", "sourceValue": 25.72362904837996, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/repair_percentile", "notes": [ "331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks" ], "uncertainty": { "level": 0.95, "low": 21.290894142129094, "high": 30.25276124764517, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "25.72362904837996", "ci_low": "21.290894142129094", "ci_high": "30.25276124764517", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "matched insert hidden-feature OOF ridge", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=alphagenome_fitted;task=repair_percentile", "source_record": "{\"ci_high\": \"30.25276124764517\", \"ci_low\": \"21.290894142129094\", \"interface\": \"matched insert hidden-feature OOF ridge\", \"model\": \"alphagenome_fitted\", \"note\": \"331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks\", \"selector\": \"model=alphagenome_fitted\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/repair-audit/table_ready_rows.csv\", \"source_sha256\": \"df33defaa2573bf74488304ac913e68ab338a9a13d818dc7a441648a3c85a44d\", \"task\": \"repair_percentile\", \"value\": \"25.72362904837996\"}", "note": "331 response targets in 110 loci; lower percentile is better; independently replayed all candidate nominations and ranks" }, "native_submission": "candidate", "native_value": 25.72362904837996 }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_embeddings_utility_log2", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "cell_selective", "value": 0.515, "displayValue": "0.515", "status": "evaluated", "interface": "fitted common+contrast, whole-locus OOF", "sourceValue": 0.5147185900778193, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/cell_selective", "notes": [ "Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family" ], "uncertainty": { "level": 0.95, "low": 0.42155293765502505, "high": 0.6117450218744611, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "0.5147185900778193", "ci_low": "0.42155293765502505", "ci_high": "0.6117450218744611", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "fitted common+contrast, whole-locus OOF", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=alphagenome_fitted;task=cell_selective", "source_record": "{\"ci_high\": \"0.6117450218744611\", \"ci_low\": \"0.42155293765502505\", \"interface\": \"fitted common+contrast, whole-locus OOF\", \"model\": \"alphagenome_fitted\", \"note\": \"Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family\", \"selector\": \"model=alphagenome_fitted\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/cell-selective/table_ready_rows.csv\", \"source_sha256\": \"40cff5ed272604bcdae3c3fd61250cd7b83f89b1ac327073004d4cd26c27403b\", \"task\": \"cell_selective\", \"value\": \"0.5147185900778193\"}", "note": "Current Fig.5C: 127 loci, 280 backgrounds, 167958 edits, lambda=1 and no-edit option; 50000 locus bootstrap; new rows retrospective pointwise95, not original24-contrast confirmatory family" }, "native_submission": "candidate", "native_value": 0.5147185900778193 }, { "configuration_id": "alphagenome_fitted", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_embeddings_gain_over_random_log2", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome_fitted", "taskId": "atlas_profile", "value": 0.833, "displayValue": "0.833", "status": "evaluated", "interface": "source-family-held-out ridge on 1536-dimensional frozen hidden representations", "sourceValue": 0.8329312309424242, "sourceId": "table-1", "sourceRecordId": "alphagenome_fitted/atlas_profile", "notes": [ "Current Fig.6A: 2411 candidates,16 overlapping programs,top10; mean gain over intent-matched random; repeat1=0.827058248061848;repeat2=0.8388042138230003; no independent-program confidence interval; distinct from native functional-output features" ], "uncertainty": null }, "archived_source_record": { "model": "alphagenome_fitted", "label": "AlphaGenome (fitted)", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.8329312309424242", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-family-held-out ridge on 1536-dimensional frozen hidden representations", "source_file": "source/alphagenome-native-completion-20260922.csv", "source_sha256": "bc256131373bc013929e05fc806ad4b17d8e35120ba9f9a27ebeafd305eef528", "selector": "model=alphagenome_fitted;task=atlas_profile", "source_record": "{\"ci_high\": \"\", \"ci_low\": \"\", \"interface\": \"source-family-held-out ridge on 1536-dimensional frozen hidden representations\", \"model\": \"alphagenome_fitted\", \"note\": \"Current Fig.6A: 2411 candidates,16 overlapping programs,top10; mean gain over intent-matched random; repeat1=0.827058248061848;repeat2=0.8388042138230003; no independent-program confidence interval; distinct from native functional-output features\", \"selector\": \"model=alphagenome_fitted\", \"source_file\": \"/Users/fvk220/Documents/New project/output/alphagenome-table-completion-20260922/atlas/results/summary.csv\", \"source_sha256\": \"6b18dbaa6ab55fe96851f20dd6291f2270fda803ebdcb80ac621886d22b528b8\", \"task\": \"atlas_profile\", \"value\": \"0.8329312309424242\"}", "note": "Current Fig.6A: 2411 candidates,16 overlapping programs,top10; mean gain over intent-matched random; repeat1=0.827058248061848;repeat2=0.8388042138230003; no independent-program confidence interval; distinct from native functional-output features" }, "native_submission": "candidate", "native_value": 0.8329312309424242 }, { "configuration_id": "alphagenome", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "prior", "value": null, "displayValue": "—", "status": "unsupported", "interface": null, "sourceValue": null, "sourceId": "table-1", "sourceRecordId": "alphagenome/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": null, "manuscriptStatus": "unsupported", "missingReason": "AlphaGenome predicts functional measurements, not nucleotide probabilities for a hidden base.", "extensionId": "coverage-20261027" }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "unsupported", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=alphagenome;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"\", \"ci_low\": \"\", \"fitting\": \"\", \"higher_is_better\": \"\", \"holdout\": \"\", \"initialization\": \"trained\", \"interval\": \"\", \"label\": \"alphagenome\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"prior\", \"metric\": \"\", \"model\": \"alphagenome\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"\", \"sampling_unit\": \"\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"\", \"source_sha256\": \"\", \"status\": \"unsupported\", \"task\": \"prior\", \"task_label\": \"\", \"value\": \"\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "omitted_missing_or_unsupported_never_zero" }, { "configuration_id": "alphagenome", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_native_outputs_spearman_rho", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "opensplice", "value": 0.621, "displayValue": "0.621", "status": "evaluated", "interface": "published minigene score", "sourceValue": 0.6205920024993516, "sourceId": "table-1", "sourceRecordId": "alphagenome/opensplice", "notes": [ "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." ], "uncertainty": { "level": 0.95, "low": 0.6048729281151268, "high": 0.6365323275300002, "type": "reported pointwise bootstrap interval", "samplingUnit": "gene", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.6205920024993516", "ci_low": "0.6048729281151268", "ci_high": "0.6365323275300002", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "published minigene score", "source_file": "provenance/opensplice-table-20260923/transfer-plotted.csv", "source_sha256": "c87c022423e7c25e996ee4abd8f4cce1920a5a396eece7056e8302a58c634c98", "selector": "panel=A;random=False;key=alphagenome", "source_record": "{\"figure\": \"transfer\", \"panel\": \"A\", \"key\": \"alphagenome\", \"label\": \"AlphaGenome \\u00b7 native\", \"value\": \"0.6205920024993516\", \"low\": \"0.6048729281151268\", \"high\": \"0.6365323275300002\", \"x\": \"\", \"role\": \"teal\", \"random\": \"False\", \"interface\": \"published minigene score\", \"population\": \"590,104 interventions / 466 genes\", \"interval\": \"95% gene-bootstrap interval\", \"source\": \"results/figure2_opensplice_added.json\", \"selector\": \"reports.alphagenome\"}", "note": "590104 interventions;466 genes; five-fold complete-gene readouts or separately identified native output, as Figure2A." }, "native_submission": "candidate", "native_value": 0.6205920024993516 }, { "configuration_id": "alphagenome", "scientific_task_id": "context", "native_protocol_id": "context_native_outputs_spearman_rho", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "context", "value": 0.398, "displayValue": "0.398", "status": "evaluated", "interface": "Native accessibility output", "sourceValue": 0.3979143979814734, "sourceId": "table-1", "sourceRecordId": "alphagenome/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.3657660158913222, "high": 0.4311824782669762, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.3979143979814734", "ci_low": "0.3657660158913222", "ci_high": "0.4311824782669762", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Native accessibility output", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=alphagenome;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"alphagenome_research_version\\\":\\\"0.3.0\\\",\\\"alphagenome_version\\\":\\\"0.8.0\\\",\\\"analysis\\\":\\\"native AlphaGenome cell-matched accessibility effects on Siraj\\\",\\\"background_hg38_mismatch_range\\\":[0,139],\\\"backgrounds\\\":284,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"credential_recorded\\\":false,\\\"device\\\":\\\"cuda:0\\\",\\\"elapsed_seconds_this_run\\\":940.1864468071144,\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"jax_version\\\":\\\"0.10.2\\\",\\\"modalities\\\":[\\\"atac\\\",\\\"dnase\\\"],\\\"model_repo\\\":\\\"google/alphagenome-all-folds\\\",\\\"model_revision\\\":\\\"a8f293a76ee73d5b57f3bf2ae146510589fcf187\\\",\\\"ontology_terms\\\":{\\\"hepg2\\\":\\\"EFO:0001187\\\",\\\"k562\\\":\\\"EFO:0002067\\\"},\\\"output_sha256\\\":\\\"8222d5db1e048ba392921baa4beeace5c5451968b8fb5f1a67d8a937b3e5feaa\\\",\\\"primary_score\\\":\\\"mean of ATAC and DNase signed difference-sums after division by each track's label-independent nonzero_mean\\\",\\\"python_version\\\":\\\"3.11.13\\\",\\\"records\\\":170292,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"research_code_commit\\\":\\\"1e55dcffb98ba26b31e74edc5e9f038f54c0e89d\\\",\\\"sequence_length\\\":131072,\\\"target_labels_read_or_used\\\":false,\\\"units\\\":128,\\\"widths\\\":[501,2001,10001]}\", \"ci_high\": \"0.4311824782669762\", \"ci_low\": \"0.3657660158913222\", \"fitting\": \"No Siraj assay readout fitting\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"AlphaGenome\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"alphagenome\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Native accessibility output\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"endpoint=context; model=alphagenome_accessibility_501\", \"simultaneous_ci_high\": \"0.4481812855025718\", \"simultaneous_ci_low\": \"0.3507433689420715\", \"source\": \"results/siraj_alphagenome_native.summary.csv\", \"source_sha256\": \"d7fed17feee353fef13ad62fdf8690ef1867b8fb70c2c1b4b408fd4f4ef47339\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.3979143979814734\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.3979143979814734 }, { "configuration_id": "alphagenome", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_native_outputs_accuracy_percent", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "tissue_identification", "value": 68.6, "displayValue": "68.6", "status": "evaluated", "interface": "native RNA outputs", "sourceValue": 68.62745098039215, "sourceId": "table-1", "sourceRecordId": "alphagenome/tissue_identification", "notes": [ "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" ], "uncertainty": null }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "68.62745098039215", "ci_low": "60.78431372549019", "ci_high": "72.54901960784314", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native RNA outputs", "source_file": "source/gtex-results.csv", "source_sha256": "ef1bc015234551618dccc3adb36e0b151ff3f39073667a443c9c7fd03463e921", "selector": "model=alphagenome;metric=retrieval_top1;scale=100", "source_record": "{\"bootstrap_replicates\": \"10000\", \"high\": \"0.7254901960784313\", \"invalid_replicates\": \"0\", \"low\": \"0.6078431372549019\", \"metric\": \"retrieval_top1\", \"model\": \"alphagenome\", \"n_genes\": \"5782\", \"n_tissues\": \"51\", \"replicates_with_constant_profiles\": \"0\", \"value\": \"0.6862745098039216\"}", "note": "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" }, "native_submission": "candidate", "native_value": 68.62745098039215 }, { "configuration_id": "alphagenome", "scientific_task_id": "background", "native_protocol_id": "background_native_outputs_spearman_rho", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "background", "value": 0.26, "displayValue": "0.260", "status": "evaluated", "interface": "Native accessibility output", "sourceValue": 0.2595833829724495, "sourceId": "table-1", "sourceRecordId": "alphagenome/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.21923266239249647, "high": 0.2993448513833979, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.2595833829724495", "ci_low": "0.21923266239249647", "ci_high": "0.2993448513833979", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Native accessibility output", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=alphagenome;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"run\\\":{\\\"alphagenome_research_version\\\":\\\"0.3.0\\\",\\\"alphagenome_version\\\":\\\"0.8.0\\\",\\\"analysis\\\":\\\"native AlphaGenome cell-matched accessibility effects on Siraj\\\",\\\"background_hg38_mismatch_range\\\":[0,139],\\\"backgrounds\\\":284,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"credential_recorded\\\":false,\\\"device\\\":\\\"cuda:0\\\",\\\"elapsed_seconds_this_run\\\":940.1864468071144,\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"jax_version\\\":\\\"0.10.2\\\",\\\"modalities\\\":[\\\"atac\\\",\\\"dnase\\\"],\\\"model_repo\\\":\\\"google/alphagenome-all-folds\\\",\\\"model_revision\\\":\\\"a8f293a76ee73d5b57f3bf2ae146510589fcf187\\\",\\\"ontology_terms\\\":{\\\"hepg2\\\":\\\"EFO:0001187\\\",\\\"k562\\\":\\\"EFO:0002067\\\"},\\\"output_sha256\\\":\\\"8222d5db1e048ba392921baa4beeace5c5451968b8fb5f1a67d8a937b3e5feaa\\\",\\\"primary_score\\\":\\\"mean of ATAC and DNase signed difference-sums after division by each track's label-independent nonzero_mean\\\",\\\"python_version\\\":\\\"3.11.13\\\",\\\"records\\\":170292,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"research_code_commit\\\":\\\"1e55dcffb98ba26b31e74edc5e9f038f54c0e89d\\\",\\\"sequence_length\\\":131072,\\\"target_labels_read_or_used\\\":false,\\\"units\\\":128,\\\"widths\\\":[501,2001,10001]}}\", \"ci_high\": \"0.2993448513833979\", \"ci_low\": \"0.21923266239249647\", \"fitting\": \"No Siraj assay readout fitting\", \"higher_is_better\": \"True\", \"holdout\": \"Same mutation on two backgrounds; 114 elements / 68,026 pairs\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"AlphaGenome\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"alphagenome\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Native accessibility output\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=alphagenome\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_background_interaction.summary.csv\", \"source_sha256\": \"13fc40cd51049436fb416d7c5c5de74d9c4b5f084be6534276d64ab3e1698c4d\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj sequence-background interaction\", \"value\": \"0.2595833829724495\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.2595833829724495 }, { "configuration_id": "alphagenome", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_native_outputs_residual_midrank_percentile", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "repair_percentile", "value": 24.3, "displayValue": "24.3", "status": "evaluated", "interface": "native assay-matched outputs", "sourceValue": 24.33125909263845, "sourceId": "table-1", "sourceRecordId": "alphagenome/repair_percentile", "notes": [ "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" ], "uncertainty": { "level": 0.95, "low": 20.474510192588795, "high": 28.25128235491262, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "24.33125909263845", "ci_low": "20.474510192588795", "ci_high": "28.25128235491262", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native assay-matched outputs", "source_file": "source/repair-summary.csv", "source_sha256": "433befcbc4b8c0d5d7b3779c2a2aa1c15257bd5a4cd28aaeb35ae0ccbd1ca5a7", "selector": "model=alphagenome;value=100*candidate_percentile", "source_record": "{\"candidate_percentile\": \"0.2433125909263845\", \"ci_high\": \"0.2825128235491262\", \"ci_low\": \"0.20474510192588796\", \"elements\": \"110\", \"family\": \"native\", \"label\": \"AlphaGenome \\u00b7 zero shot\", \"model\": \"alphagenome\", \"resolved_instances\": \"331\", \"top1_success_ci_high\": \"0.29318181818181815\", \"top1_success_ci_low\": \"0.18558712121212118\", \"top1_success_rate\": \"0.23712121212121212\"}", "note": "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" }, "native_submission": "candidate", "native_value": 24.33125909263845 }, { "configuration_id": "alphagenome", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_native_outputs_utility_log2", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "cell_selective", "value": 0.384, "displayValue": "0.384", "status": "evaluated", "interface": "native outputs with one held-locus nonnegative slope per cell", "sourceValue": 0.3844643285391621, "sourceId": "table-1", "sourceRecordId": "alphagenome/cell_selective", "notes": [ "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" ], "uncertainty": { "level": 0.95, "low": 0.28837056320631366, "high": 0.483381753986141, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "0.3844643285391621", "ci_low": "0.28837056320631366", "ci_high": "0.483381753986141", "simultaneous_ci_low": "0.23732604241310754", "simultaneous_ci_high": "0.5434185232987693", "interface": "native outputs with one held-locus nonnegative slope per cell", "source_file": "source/onebase/summary.csv", "source_sha256": "9dc5c23c5346f74df63d746469c1258558e043e67b9db877c906c42733252bc6", "selector": "stream=pooled;model=alphagenome;policy=lambda_1;metric=utility", "source_record": "{\"high95\": \"0.483381753986141\", \"loci\": \"127\", \"low95\": \"0.28837056320631366\", \"mean\": \"0.3844643285391621\", \"metric\": \"utility\", \"model\": \"alphagenome\", \"policy\": \"lambda_1\", \"stream\": \"pooled\"}", "note": "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" }, "native_submission": "candidate", "native_value": 0.3844643285391621 }, { "configuration_id": "alphagenome", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_native_outputs_gain_over_random_log2", "checkpoint_target": "google/alphagenome-all-folds", "measurement": { "modelId": "alphagenome", "taskId": "atlas_profile", "value": 0.707, "displayValue": "0.707", "status": "evaluated", "interface": "source-family-held-out ridge on frozen functional features", "sourceValue": 0.7073047054037112, "sourceId": "table-1", "sourceRecordId": "alphagenome/atlas_profile", "notes": [ "16 overlapping programs in 12 cells; mean utility gain over intent-matched random of top10 candidates; mean over two measurement repeats; U=minON-max(0,maxOFF); no independent-program confidence interval; two-repeat range is descriptive; Enformer uses corrected TensorFlow-gamma extraction and refitting", "Table 1 ‡: a linear head is fitted to native functional outputs; this endpoint is not a direct zero-shot native prediction." ], "uncertainty": null }, "archived_source_record": { "model": "alphagenome", "label": "AlphaGenome", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.7073047054037112", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-family-held-out ridge on frozen functional features", "source_file": "source/atlas-macro.csv", "source_sha256": "0aff9477bb624029b2b6856a2205e3f9c88a4d70f6fdaa51f75da088e32ecf7c", "selector": "model=alphagenome;selector=full_profile;k=10;value=held_u_minus_intended_random", "source_record": "{\"held_capped_max_off\": \"1.3866212817445893\", \"held_max_off\": \"1.3851200093950362\", \"held_min_on\": \"1.1650983915242266\", \"held_u\": \"-0.2215228902203627\", \"held_u_minus_intended_random\": \"0.7073047054037112\", \"intended_random_success\": \"0.05729323275524626\", \"intended_random_u\": \"-0.9288275956240738\", \"k\": \"10\", \"model\": \"alphagenome\", \"selector\": \"full_profile\", \"success_minus_intended_random\": \"0.10833176724475374\", \"success_u_ge1\": \"0.165625\"}", "note": "16 overlapping programs in 12 cells; mean utility gain over intent-matched random of top10 candidates; mean over two measurement repeats; U=minON-max(0,maxOFF); no independent-program confidence interval; two-repeat range is descriptive; Enformer uses corrected TensorFlow-gamma extraction and refitting" }, "native_submission": "candidate", "native_value": 0.7073047054037112 }, { "configuration_id": "enformer", "scientific_task_id": "prior", "native_protocol_id": "prior_loss_bits_per_base", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "prior", "value": null, "displayValue": "—", "status": "unsupported", "interface": null, "sourceValue": null, "sourceId": "table-1", "sourceRecordId": "enformer/prior", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": null, "manuscriptStatus": "unsupported", "missingReason": "Enformer predicts functional measurements, not nucleotide probabilities for a hidden base.", "extensionId": "coverage-20261027" }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "prior", "category": "Sequence priors", "metric": "bits", "higher_is_better": "False", "status": "unsupported", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=enformer;main_table_column=prior", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"\", \"ci_high\": \"\", \"ci_low\": \"\", \"fitting\": \"\", \"higher_is_better\": \"\", \"holdout\": \"\", \"initialization\": \"trained\", \"interval\": \"\", \"label\": \"enformer\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"prior\", \"metric\": \"\", \"model\": \"enformer\", \"n_decisions\": \"\", \"n_records\": \"\", \"n_units\": \"\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"\", \"sampling_unit\": \"\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"\", \"source_sha256\": \"\", \"status\": \"unsupported\", \"task\": \"prior\", \"task_label\": \"\", \"value\": \"\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "omitted_missing_or_unsupported_never_zero" }, { "configuration_id": "enformer", "scientific_task_id": "opensplice", "native_protocol_id": "opensplice_native_outputs_spearman_rho", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "opensplice", "value": null, "displayValue": "—", "status": "unsupported", "interface": null, "sourceValue": null, "sourceId": "table-1", "sourceRecordId": "enformer/opensplice", "notes": [ "No matching full OpenSplice gene-held-out result available; transfer scores not substituted." ], "uncertainty": null, "manuscriptStatus": "unsupported", "missingReason": "Enformer has no native splice output. A splicing predictor fitted to its representations would be a separate model/readout entry.", "extensionId": "coverage-20261027" }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "opensplice", "category": "Functional prediction", "metric": "macro-gene Spearman", "higher_is_better": "True", "status": "unsupported", "value": "", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "", "source_file": "", "source_sha256": "", "selector": "", "source_record": "", "note": "No matching full OpenSplice gene-held-out result available; transfer scores not substituted." }, "native_submission": "omitted_missing_or_unsupported_never_zero" }, { "configuration_id": "enformer", "scientific_task_id": "context", "native_protocol_id": "context_native_outputs_spearman_rho", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "context", "value": 0.282, "displayValue": "0.282", "status": "evaluated", "interface": "Native accessibility output", "sourceValue": 0.2817834625224349, "sourceId": "table-1", "sourceRecordId": "enformer/context", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.2423146232563836, "high": 0.3212672337418212, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "context", "category": "Biological context", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.2817834625224349", "ci_low": "0.2423146232563836", "ci_high": "0.3212672337418212", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Native accessibility output", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=enformer;main_table_column=context", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"canonical\\\":{\\\"path\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\"},\\\"merged_cache\\\":{\\\"path\\\":\\\"data/cache/enformer_siraj_dnase.parquet\\\",\\\"sha256\\\":\\\"796ccff1b070ffa5043c3a2283ef6be4d8fd26dd5448a0ac61624fc3c07f8097\\\"},\\\"shards\\\":[{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":594.9561964508612,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"a4bedb5b9c9446f2755b751a7f461ea8713e8fbd61a0812a0e56ffbf6bef218b\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42565,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":0,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard0.parquet\\\",\\\"sha256\\\":\\\"a4bedb5b9c9446f2755b751a7f461ea8713e8fbd61a0812a0e56ffbf6bef218b\\\"},{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":594.9562270110473,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"4140e012b651ca7757cda84cccd00ffe2ffa2754df88d4bbb2aea3f8652e054d\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42574,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":1,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard1.parquet\\\",\\\"sha256\\\":\\\"4140e012b651ca7757cda84cccd00ffe2ffa2754df88d4bbb2aea3f8652e054d\\\"},{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":5.483312827069312,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"35d79c27bb6b631c1249c6f5ea149ed0e58cbc5e8e0f5c8f7ea2d9f2f14dc0f3\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42578,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":2,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard2.parquet\\\",\\\"sha256\\\":\\\"35d79c27bb6b631c1249c6f5ea149ed0e58cbc5e8e0f5c8f7ea2d9f2f14dc0f3\\\"},{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":5.483222619863227,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"d27712a262c25ba8405525f7e79d912fd8ade6f12bffe4969fe8754042133c26\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42575,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":3,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard3.parquet\\\",\\\"sha256\\\":\\\"d27712a262c25ba8405525f7e79d912fd8ade6f12bffe4969fe8754042133c26\\\"}]}\", \"ci_high\": \"0.3212672337418212\", \"ci_low\": \"0.2423146232563836\", \"fitting\": \"No Siraj assay readout fitting\", \"higher_is_better\": \"True\", \"holdout\": \"Complete held-out elements; fixed HepG2-minus-K562 contrast; 128 / 170,292\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"Enformer\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"context\", \"metric\": \"macro-element Spearman\", \"model\": \"enformer\", \"n_decisions\": \"\", \"n_records\": \"170292\", \"n_units\": \"128\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Native accessibility output\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"endpoint=context; model=enformer_dnase_1024\", \"simultaneous_ci_high\": \"0.3437482704874167\", \"simultaneous_ci_low\": \"0.2174394039751466\", \"source\": \"results/siraj_alphagenome_native.summary.csv\", \"source_sha256\": \"d7fed17feee353fef13ad62fdf8690ef1867b8fb70c2c1b4b408fd4f4ef47339\", \"status\": \"evaluated\", \"task\": \"context\", \"task_label\": \"Siraj cell-context interaction\", \"value\": \"0.2817834625224349\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.2817834625224349 }, { "configuration_id": "enformer", "scientific_task_id": "tissue_identification", "native_protocol_id": "tissue_identification_native_outputs_accuracy_percent", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "tissue_identification", "value": 31.4, "displayValue": "31.4", "status": "evaluated", "interface": "native RNA outputs", "sourceValue": 31.372549019607842, "sourceId": "table-1", "sourceRecordId": "enformer/tissue_identification", "notes": [ "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" ], "uncertainty": null }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "tissue_identification", "category": "Biological context", "metric": "top-1 identification percent", "higher_is_better": "True", "status": "evaluated", "value": "31.372549019607842", "ci_low": "25.49019607843137", "ci_high": "37.254901960784316", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native RNA outputs", "source_file": "source/gtex-results.csv", "source_sha256": "ef1bc015234551618dccc3adb36e0b151ff3f39073667a443c9c7fd03463e921", "selector": "model=enformer;metric=retrieval_top1;scale=100", "source_record": "{\"bootstrap_replicates\": \"10000\", \"high\": \"0.37254901960784315\", \"invalid_replicates\": \"0\", \"low\": \"0.2549019607843137\", \"metric\": \"retrieval_top1\", \"model\": \"enformer\", \"n_genes\": \"5782\", \"n_tissues\": \"51\", \"replicates_with_constant_profiles\": \"0\", \"value\": \"0.3137254901960784\"}", "note": "51 represented tissues; 5782 common held-out genes; paired whole-gene bootstrap with reranking; fixed tissue set; no unseen-tissue transfer; chance 100/51 percent" }, "native_submission": "candidate", "native_value": 31.372549019607842 }, { "configuration_id": "enformer", "scientific_task_id": "background", "native_protocol_id": "background_native_outputs_spearman_rho", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "background", "value": 0.181, "displayValue": "0.181", "status": "evaluated", "interface": "Native accessibility output", "sourceValue": 0.1811050695712191, "sourceId": "table-1", "sourceRecordId": "enformer/background", "notes": [ "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" ], "uncertainty": { "level": 0.95, "low": 0.14146192745254052, "high": 0.2202278305871842, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "background", "category": "Genetic interactions", "metric": "macro-element Spearman", "higher_is_better": "True", "status": "evaluated", "value": "0.1811050695712191", "ci_low": "0.14146192745254052", "ci_high": "0.2202278305871842", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "Native accessibility output", "source_file": "source/manuscript-reference.csv", "source_sha256": "c46be05c3a3a0e59047c6596b1d1275949989903cef4dd3bc9618cc8450f5ba2", "selector": "model=enformer;main_table_column=background", "source_record": "{\"bend2_main_block\": \"prediction\", \"budget\": \"\", \"budget_unit\": \"\", \"checkpoint_metadata\": \"{\\\"run\\\":{\\\"canonical\\\":{\\\"path\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\"},\\\"merged_cache\\\":{\\\"path\\\":\\\"data/cache/enformer_siraj_dnase.parquet\\\",\\\"sha256\\\":\\\"796ccff1b070ffa5043c3a2283ef6be4d8fd26dd5448a0ac61624fc3c07f8097\\\"},\\\"shards\\\":[{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":594.9561964508612,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"a4bedb5b9c9446f2755b751a7f461ea8713e8fbd61a0812a0e56ffbf6bef218b\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42565,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":0,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard0.parquet\\\",\\\"sha256\\\":\\\"a4bedb5b9c9446f2755b751a7f461ea8713e8fbd61a0812a0e56ffbf6bef218b\\\"},{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":594.9562270110473,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"4140e012b651ca7757cda84cccd00ffe2ffa2754df88d4bbb2aea3f8652e054d\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42574,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":1,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard1.parquet\\\",\\\"sha256\\\":\\\"4140e012b651ca7757cda84cccd00ffe2ffa2754df88d4bbb2aea3f8652e054d\\\"},{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":5.483312827069312,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"35d79c27bb6b631c1249c6f5ea149ed0e58cbc5e8e0f5c8f7ea2d9f2f14dc0f3\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42578,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":2,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard2.parquet\\\",\\\"sha256\\\":\\\"35d79c27bb6b631c1249c6f5ea149ed0e58cbc5e8e0f5c8f7ea2d9f2f14dc0f3\\\"},{\\\"metadata\\\":{\\\"ambiguous_reference_encoding\\\":\\\"N is encoded as Enformer's fifth index, which becomes an all-zero A/C/G/T vector in enformer-pytorch\\\",\\\"analysis\\\":\\\"native Enformer cell-matched DNase effects on Siraj\\\",\\\"backgrounds\\\":71,\\\"chain\\\":\\\"data/raw/gpn_msa/hg19ToHg38.over.chain.gz\\\",\\\"chain_sha256\\\":\\\"5c0598e500ceb5a78c73086929e8ef993aec309bcafb595139b53d440b125a1d\\\",\\\"checkpoint_sha256\\\":\\\"99b09d602e195d89c7d4debe144bb2f43907ba0d74006e97098e99d9171c439c\\\",\\\"device\\\":\\\"cuda\\\",\\\"device_name\\\":\\\"NVIDIA GH200 120GB\\\",\\\"elapsed_seconds_this_run\\\":5.483222619863227,\\\"enformer_pytorch_version\\\":\\\"0.8.8\\\",\\\"experimental_haplotype_embedded\\\":true,\\\"input\\\":\\\"data/processed/siraj_satmut/canonical.parquet\\\",\\\"input_sha256\\\":\\\"2c77d02359abdc602526b17b953d3a26fd2b88fe3ecb7f04e1fea3874c1432e8\\\",\\\"local_files_only\\\":true,\\\"model\\\":\\\"EleutherAI/enformer-official-rough\\\",\\\"model_parameter_count_before_task_slice\\\":251221292,\\\"num_shards\\\":4,\\\"output_resolution\\\":128,\\\"output_sha256\\\":\\\"d27712a262c25ba8405525f7e79d912fd8ade6f12bffe4969fe8754042133c26\\\",\\\"primary_width\\\":1024,\\\"python_version\\\":\\\"3.13.5\\\",\\\"records\\\":42575,\\\"reference\\\":\\\"data/raw/hg38_twobit/hg38.2bit\\\",\\\"reference_sha256\\\":\\\"1f67aaa17a77b327738fe750ab37430a85dddf73c7d9189385ad1839259acec0\\\",\\\"revision\\\":\\\"affe5713ae9017460706a44108289b13c5fee16c\\\",\\\"selected_track_descriptions\\\":{\\\"hepg2\\\":\\\"DNASE:HepG2\\\",\\\"k562\\\":\\\"DNASE:K562\\\"},\\\"selected_tracks\\\":{\\\"hepg2\\\":[27,91,234],\\\"k562\\\":[121,122,123,625]},\\\"sequence_length\\\":196608,\\\"shard_index\\\":3,\\\"target_labels_read_or_used\\\":false,\\\"targets\\\":\\\"data/raw/enformer_targets_human/targets_human.txt\\\",\\\"targets_sha256\\\":\\\"d90233175a9fec389e5ac04cc4aac8cfc08c4aa86217e2e5c10221de78023580\\\",\\\"torch_version\\\":\\\"2.13.0+cu130\\\",\\\"units\\\":71,\\\"widths\\\":[1024,4096,16384]},\\\"path\\\":\\\"/e/scratch/e-dev-2026d02-064/gimbench_enformer_siraj/enformer_siraj.shard3.parquet\\\",\\\"sha256\\\":\\\"d27712a262c25ba8405525f7e79d912fd8ade6f12bffe4969fe8754042133c26\\\"}]}}\", \"ci_high\": \"0.2202278305871842\", \"ci_low\": \"0.14146192745254052\", \"fitting\": \"No Siraj assay readout fitting\", \"higher_is_better\": \"True\", \"holdout\": \"Same mutation on two backgrounds; 114 elements / 68,026 pairs\", \"initialization\": \"trained\", \"interval\": \"marginal 95% biological-unit bootstrap\", \"label\": \"Enformer\", \"library\": \"\", \"main_readout\": \"\", \"main_table_column\": \"background\", \"metric\": \"macro-element Spearman\", \"model\": \"enformer\", \"n_decisions\": \"\", \"n_records\": \"68026\", \"n_units\": \"114\", \"notes\": \"\", \"overlay_protocol\": \"\", \"overlay_selector\": \"\", \"overlay_source\": \"\", \"overlay_source_sha256\": \"\", \"policy\": \"\", \"query\": \"\", \"readout\": \"Native accessibility output\", \"sampling_unit\": \"element\", \"seed_max\": \"\", \"seed_min\": \"\", \"seed_sd\": \"\", \"selector\": \"model=enformer\", \"simultaneous_ci_high\": \"\", \"simultaneous_ci_low\": \"\", \"source\": \"results/siraj_background_interaction.summary.csv\", \"source_sha256\": \"13fc40cd51049436fb416d7c5c5de74d9c4b5f084be6534276d64ab3e1698c4d\", \"status\": \"evaluated\", \"task\": \"background\", \"task_label\": \"Siraj sequence-background interaction\", \"value\": \"0.1811050695712191\", \"variant\": \"primary\"}", "note": "Preserved current manuscript's selected interface; upstream source and interval definitions in source_record" }, "native_submission": "candidate", "native_value": 0.1811050695712191 }, { "configuration_id": "enformer", "scientific_task_id": "repair_percentile", "native_protocol_id": "repair_percentile_native_outputs_residual_midrank_percentile", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "repair_percentile", "value": 31.5, "displayValue": "31.5", "status": "evaluated", "interface": "native assay-matched outputs", "sourceValue": 31.520239012563213, "sourceId": "table-1", "sourceRecordId": "enformer/repair_percentile", "notes": [ "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" ], "uncertainty": { "level": 0.95, "low": 27.445455234719134, "high": 35.664422498636306, "type": "reported pointwise bootstrap interval", "samplingUnit": "element", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "repair_percentile", "category": "Design decisions", "metric": "measured rank percentile", "higher_is_better": "False", "status": "evaluated", "value": "31.520239012563213", "ci_low": "27.445455234719134", "ci_high": "35.664422498636306", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "native assay-matched outputs", "source_file": "source/repair-summary.csv", "source_sha256": "433befcbc4b8c0d5d7b3779c2a2aa1c15257bd5a4cd28aaeb35ae0ccbd1ca5a7", "selector": "model=enformer;value=100*candidate_percentile", "source_record": "{\"candidate_percentile\": \"0.3152023901256321\", \"ci_high\": \"0.35664422498636306\", \"ci_low\": \"0.27445455234719135\", \"elements\": \"110\", \"family\": \"native\", \"label\": \"Enformer \\u00b7 zero shot\", \"model\": \"enformer\", \"resolved_instances\": \"331\", \"top1_success_ci_high\": \"0.2409090909090909\", \"top1_success_ci_low\": \"0.1409090909090909\", \"top1_success_rate\": \"0.18939393939393942\"}", "note": "331 response targets in 110 elements; measured residual midrank scaled to [0,100], lower better; first nomination minimizes model-space distance to its predicted direct-swap effect; no direct reversion" }, "native_submission": "candidate", "native_value": 31.520239012563213 }, { "configuration_id": "enformer", "scientific_task_id": "cell_selective", "native_protocol_id": "cell_selective_native_outputs_utility_log2", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "cell_selective", "value": 0.29, "displayValue": "0.290", "status": "evaluated", "interface": "native outputs with one held-locus nonnegative slope per cell", "sourceValue": 0.29029819492545156, "sourceId": "table-1", "sourceRecordId": "enformer/cell_selective", "notes": [ "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" ], "uncertainty": { "level": 0.95, "low": 0.19281936676828326, "high": 0.39092928522050263, "type": "reported pointwise bootstrap interval", "samplingUnit": "locus", "sourceId": "table-1-records", "note": "Preserved from the matched source-table record; not an interval for the derived reference index and not evidence for a between-model rank." } }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "cell_selective", "category": "Design decisions", "metric": "log2 reporter utility", "higher_is_better": "True", "status": "evaluated", "value": "0.29029819492545156", "ci_low": "0.19281936676828326", "ci_high": "0.39092928522050263", "simultaneous_ci_low": "0.13445653355830464", "simultaneous_ci_high": "0.44630921373767996", "interface": "native outputs with one held-locus nonnegative slope per cell", "source_file": "source/onebase/summary.csv", "source_sha256": "9dc5c23c5346f74df63d746469c1258558e043e67b9db877c906c42733252bc6", "selector": "stream=pooled;model=enformer;policy=lambda_1;metric=utility", "source_record": "{\"high95\": \"0.39092928522050263\", \"loci\": \"127\", \"low95\": \"0.19281936676828326\", \"mean\": \"0.29029819492545156\", \"metric\": \"utility\", \"model\": \"enformer\", \"policy\": \"lambda_1\", \"stream\": \"pooled\"}", "note": "127 equal-weight loci; both cell requests and parental backgrounds averaged within locus; U=target minus absolute other-cell effect, lambda=1; no edit available; pointwise and original 24-contrast simultaneous intervals retained" }, "native_submission": "candidate", "native_value": 0.29029819492545156 }, { "configuration_id": "enformer", "scientific_task_id": "atlas_profile", "native_protocol_id": "atlas_profile_native_outputs_gain_over_random_log2", "checkpoint_target": "EleutherAI/enformer-official-rough", "measurement": { "modelId": "enformer", "taskId": "atlas_profile", "value": 0.7, "displayValue": "0.700", "status": "evaluated", "interface": "source-family-held-out ridge on frozen functional features", "sourceValue": 0.7004804914893465, "sourceId": "table-1", "sourceRecordId": "enformer/atlas_profile", "notes": [ "16 overlapping programs in 12 cells; mean utility gain over intent-matched random of top10 candidates; mean over two measurement repeats; U=minON-max(0,maxOFF); no independent-program confidence interval; two-repeat range is descriptive; Enformer uses corrected TensorFlow-gamma extraction and refitting", "Table 1 ‡: a linear head is fitted to native functional outputs; this endpoint is not a direct zero-shot native prediction." ], "uncertainty": null }, "archived_source_record": { "model": "enformer", "label": "Enformer", "task": "atlas_profile", "category": "Design decisions", "metric": "log2 reporter utility gain over intent-matched random", "higher_is_better": "True", "status": "evaluated", "value": "0.7004804914893465", "ci_low": "", "ci_high": "", "simultaneous_ci_low": "", "simultaneous_ci_high": "", "interface": "source-family-held-out ridge on frozen functional features", "source_file": "source/atlas-macro.csv", "source_sha256": "0aff9477bb624029b2b6856a2205e3f9c88a4d70f6fdaa51f75da088e32ecf7c", "selector": "model=enformer;selector=full_profile;k=10;value=held_u_minus_intended_random", "source_record": "{\"held_capped_max_off\": \"1.4020741284572176\", \"held_max_off\": \"1.4016009932172384\", \"held_min_on\": \"1.1737270243224902\", \"held_u\": \"-0.22834710413472722\", \"held_u_minus_intended_random\": \"0.7004804914893465\", \"intended_random_success\": \"0.05729323275524626\", \"intended_random_u\": \"-0.9288275956240738\", \"k\": \"10\", \"model\": \"enformer\", \"selector\": \"full_profile\", \"success_minus_intended_random\": \"0.10520676724475374\", \"success_u_ge1\": \"0.1625\"}", "note": "16 overlapping programs in 12 cells; mean utility gain over intent-matched random of top10 candidates; mean over two measurement repeats; U=minON-max(0,maxOFF); no independent-program confidence interval; two-repeat range is descriptive; Enformer uses corrected TensorFlow-gamma extraction and refitting" }, "native_submission": "candidate", "native_value": 0.7004804914893465 } ], "other_rows_retained_in_source_snapshot": [ { "id": "markov0", "label": "Base composition", "shortLabel": "Base composition", "family": "Composition", "kind": "baseline", "interface": "sequence", "parentModelId": null, "parameterLabel": null, "description": "Empirical base-composition predictor; no assay labels.", "sortOrder": 0 }, { "id": "kmer_3_6", "label": "k-mer ridge", "shortLabel": "k-mer ridge", "family": "k-mer", "kind": "baseline", "interface": "fitted", "parentModelId": null, "parameterLabel": null, "description": "Ridge regression on 3–6-mer features; multi-cell uses strand-invariant 1–6-mer frequencies.", "sortOrder": 14 }, { "id": "onehot_resnet", "label": "Task-trained ResNet", "shortLabel": "ResNet", "family": "ResNet", "kind": "baseline", "interface": "task-trained", "parentModelId": null, "parameterLabel": null, "description": "Supervised one-hot network trained on the task source data; multi-cell uses a three-seed ensemble.", "sortOrder": 15 }, { "id": "spliceai", "label": "SpliceAI", "shortLabel": "SpliceAI", "family": "SpliceAI", "kind": "specialist", "interface": "native", "parentModelId": null, "parameterLabel": null, "description": "Specialist splice predictor, evaluated using its published minigene score; historical supervision differs from fitted benchmark readouts.", "sortOrder": 16 } ], "shared_prior_deduplications": [ { "repository": "InstaDeepAI/NTv3_100M_post", "task_id": "prior_loss_bits_per_base", "configuration_id": "ntv3_100m_native", "value": 1.492461935201394 }, { "repository": "InstaDeepAI/NTv3_650M_post", "task_id": "prior_loss_bits_per_base", "configuration_id": "ntv3_650m_native", "value": 1.3745521953833992 } ], "counts": { "scientific_tasks": 8, "native_protocol_ids": 15, "foundation_readout_configurations": 13, "checkpoint_repositories": 10, "source_cells": 104, "evaluated_source_cells": 99, "native_candidate_results_after_shared_prior_deduplication": 97, "missing_source_cells": 5 } }