Register BEND 2.0 genomic benchmark protocols and provenance

Eight scientific tasks use fifteen readout-specific subleaderboards. Published metrics, units and directions are preserved; dataset data and README are unchanged. Native leaderboard activation requires Hugging Face beta allow-list approval.

SHA256SUMS

@@ -0,0 +1,9 @@
+# SHA-256 for the BEND 2.0 native benchmark registration files only.
+be2875eed0a47dbb99853244c3f0a69551d5ca0c5a2ec4e28bc4e93915363819  benchmark-results.json
+073a35d6c92ee005487e67ac0703a49a1cea8886b8d7c19c3c4def24c3a6ea76  eval.yaml
+f0d3657727953cf55f47c038099d75ed861b019cfd87d22314786d6f712a8b9d  source-snapshot/data/benchmark.json
+02a6a380db3093b8a6d75faafc094e65539cecc4ef604cd55dbefa89609c347b  source-snapshot/data/evaluation-updates.json
+332d38d328592db946c336c0260d62c4aa02c9be005636e29bea5054fa743c28  source-snapshot/provenance/table-1-source-records.csv
+75febce7f238262b435731ba8258f7f647111a6edf83f430f2099ada3eb5c471  source-snapshot/provenance/table-1.tex
+aa38b99913952ad9f61f5cc7e0df06b1557b5e5bc8d8cec6d83a552ee18d142c  task-mapping.md
+ef9f5b46451d0b8f3a65725d242300d73d0b2f39e9177b8aed6707bbae37e642  validation.json

eval.yaml

@@ -0,0 +1,35 @@
+# BEND 2.0 registration: 8 scientific tasks, 15 protocol subleaderboards.
+# Only documented fields and the HF-maintained lower_is_better convention are used.
+# BEND 2.0 is not yet in the supported framework enum, so evaluation_framework is omitted.
+# config/split are omitted: packaged mixed-format cross-validation is authoritative.
+# Metric units are encoded in task IDs; task-mapping.md defines each protocol.
+name: BEND 2.0
+description: 'Benchmarking the capabilities needed to design DNA: eight scientific
+  tasks, with separate frozen-embedding and native-output protocols on seven downstream
+  tasks (15 subleaderboards including one shared sequence-prior endpoint). Metrics
+  retain their published units and direction. Native-output multi-cell design still
+  fits a source-only linear head; cell-selective native outputs use source-only calibration.
+  Protocols and complete model/readout comparisons: https://machinelearninglifescience.github.io/BEND2/.
+  Use the BEND 2.0 package and task-specific held-gene, held-locus or held-family
+  protocols; Hub viewer data labels are not evaluation splits.'
+leaderboard:
+  base_models_only: false
+tasks:
+- id: prior_loss_bits_per_base
+  lower_is_better: true
+- id: opensplice_embeddings_spearman_rho
+- id: opensplice_native_outputs_spearman_rho
+- id: context_embeddings_spearman_rho
+- id: context_native_outputs_spearman_rho
+- id: tissue_identification_embeddings_accuracy_percent
+- id: tissue_identification_native_outputs_accuracy_percent
+- id: background_embeddings_spearman_rho
+- id: background_native_outputs_spearman_rho
+- id: repair_percentile_embeddings_residual_midrank_percentile
+  lower_is_better: true
+- id: repair_percentile_native_outputs_residual_midrank_percentile
+  lower_is_better: true
+- id: cell_selective_embeddings_utility_log2
+- id: cell_selective_native_outputs_utility_log2
+- id: atlas_profile_embeddings_gain_over_random_log2
+- id: atlas_profile_native_outputs_gain_over_random_log2

source-snapshot/data/evaluation-updates.json

@@ -0,0 +1,209 @@
+{
+  "id": "coverage-20261027",
+  "checkedOn": "2026-11-02",
+  "sharedPriors": [
+    {
+      "modelId": "ntv3_100m_native",
+      "fromModelId": "ntv3_100m_post",
+      "checkpoint": "InstaDeepAI/NTv3_100M_post",
+      "revision": "b7292f0bd1b5b004d28561783a1890e1ec6f94fd"
+    },
+    {
+      "modelId": "ntv3_650m_native",
+      "fromModelId": "ntv3_650m_post",
+      "checkpoint": "InstaDeepAI/NTv3_650M_post",
+      "revision": "ad622051abbe9376bdea3f2727863fa559957e9f"
+    }
+  ],
+  "missing": [
+    {
+      "modelId": "alphagenome_fitted",
+      "taskId": "prior",
+      "status": "unsupported",
+      "reason": "AlphaGenome does not provide nucleotide probabilities for this hidden-base task. Fitting its functional readouts does not supply a sequence-prior head."
+    },
+    {
+      "modelId": "alphagenome",
+      "taskId": "prior",
+      "status": "unsupported",
+      "reason": "AlphaGenome predicts functional measurements, not nucleotide probabilities for a hidden base."
+    },
+    {
+      "modelId": "enformer",
+      "taskId": "prior",
+      "status": "unsupported",
+      "reason": "Enformer predicts functional measurements, not nucleotide probabilities for a hidden base."
+    },
+    {
+      "modelId": "enformer",
+      "taskId": "opensplice",
+      "status": "unsupported",
+      "reason": "Enformer has no native splice output. A splicing predictor fitted to its representations would be a separate model/readout entry."
+    },
+    {
+      "modelId": "dnabert2",
+      "taskId": "prior",
+      "status": "unsupported",
+      "reason": "DNABERT-2 predicts multi-base tokens. A forced single-base query was tested, but its tokenization mismatch prevents a directly comparable prior score.",
+      "diagnosticSourceId": "dnabert2-prior-diagnostic"
+    },
+    {
+      "modelId": "markov0",
+      "taskId": "opensplice",
+      "status": "unsupported",
+      "reason": "Base composition is a sequence-prior baseline; it does not predict functional outcomes."
+    },
+    {
+      "modelId": "markov0",
+      "taskId": "context",
+      "status": "unsupported",
+      "reason": "Base composition is a sequence-prior baseline; it does not predict functional outcomes."
+    },
+    {
+      "modelId": "markov0",
+      "taskId": "tissue_identification",
+      "status": "unsupported",
+      "reason": "Base composition is a sequence-prior baseline; it does not predict functional outcomes."
+    },
+    {
+      "modelId": "markov0",
+      "taskId": "background",
+      "status": "unsupported",
+      "reason": "Base composition is a sequence-prior baseline; it does not predict functional outcomes."
+    },
+    {
+      "modelId": "markov0",
+      "taskId": "repair_percentile",
+      "status": "unsupported",
+      "reason": "Base composition is a sequence-prior baseline; it does not predict functional outcomes."
+    },
+    {
+      "modelId": "markov0",
+      "taskId": "cell_selective",
+      "status": "unsupported",
+      "reason": "Base composition is a sequence-prior baseline; it does not predict functional outcomes."
+    },
+    {
+      "modelId": "markov0",
+      "taskId": "atlas_profile",
+      "status": "unsupported",
+      "reason": "Base composition is a sequence-prior baseline; it does not predict functional outcomes."
+    },
+    {
+      "modelId": "kmer_3_6",
+      "taskId": "prior",
+      "status": "unsupported",
+      "reason": "This baseline was trained to predict functional measurements and has no hidden-base probability output."
+    },
+    {
+      "modelId": "onehot_resnet",
+      "taskId": "prior",
+      "status": "unsupported",
+      "reason": "This baseline was trained to predict functional measurements and has no hidden-base probability output."
+    },
+    {
+      "modelId": "spliceai",
+      "taskId": "prior",
+      "status": "unsupported",
+      "reason": "SpliceAI predicts splice sites, not nucleotide probabilities for a hidden base."
+    },
+    {
+      "modelId": "spliceai",
+      "taskId": "context",
+      "status": "unsupported",
+      "reason": "SpliceAI is a splicing specialist; this endpoint requires a different functional output."
+    },
+    {
+      "modelId": "spliceai",
+      "taskId": "tissue_identification",
+      "status": "unsupported",
+      "reason": "SpliceAI is a splicing specialist; this endpoint requires a different functional output."
+    },
+    {
+      "modelId": "spliceai",
+      "taskId": "background",
+      "status": "unsupported",
+      "reason": "SpliceAI is a splicing specialist; this endpoint requires a different functional output."
+    },
+    {
+      "modelId": "spliceai",
+      "taskId": "repair_percentile",
+      "status": "unsupported",
+      "reason": "SpliceAI is a splicing specialist; this endpoint requires a different functional output."
+    },
+    {
+      "modelId": "spliceai",
+      "taskId": "cell_selective",
+      "status": "unsupported",
+      "reason": "SpliceAI is a splicing specialist; this endpoint requires a different functional output."
+    },
+    {
+      "modelId": "spliceai",
+      "taskId": "atlas_profile",
+      "status": "unsupported",
+      "reason": "SpliceAI is a splicing specialist; this endpoint requires a different functional output."
+    }
+  ],
+  "additionalSources": [
+    {
+      "id": "dnabert2-prior-diagnostic",
+      "title": "DNABERT-2 single-base query audit (excluded from scores)",
+      "snapshotPath": "provenance/coverage-20261027/dnabert2-prior/README.md",
+      "sha256": "7880ca30ea6bdbbf1cf71c78202e4e3109b546b6910ffa2a535f94439539d7b1"
+    },
+    {
+      "id": "dnabert2-prior-diagnostic-files",
+      "title": "DNABERT-2 diagnostic artifact checksums",
+      "snapshotPath": "provenance/coverage-20261027/dnabert2-prior/SHA256SUMS.json",
+      "sha256": "9e7c1cdec20b1c037ba0405d3b17a9f5daa083553ed75afa58066292856ed2d9"
+    },
+    {
+      "id": "ceno-full-opensplice",
+      "title": "CENO full OpenSplice evaluation",
+      "snapshotPath": "provenance/coverage-20261027/ceno-opensplice/entry.json",
+      "sha256": "489a42519dd6e9267ce1eff84cba8c25c9bb5feb863b15e44996cf2d5bdc7d11"
+    },
+    {
+      "id": "ceno-full-opensplice-methods",
+      "title": "CENO evaluation protocol and numerical checks",
+      "snapshotPath": "provenance/coverage-20261027/ceno-opensplice/README.md",
+      "sha256": "d0a373d57344d76116720fd8bbbc4d946ebddc33e26c5b440a5d7b1517d46e7a"
+    },
+    {
+      "id": "ceno-full-opensplice-files",
+      "title": "CENO evaluation artifact checksums",
+      "snapshotPath": "provenance/coverage-20261027/ceno-opensplice/SHA256SUMS.json",
+      "sha256": "4458b87b47a46dd8d0f678c2ea99b8151b497abf0b567c996e50c892498cac73"
+    },
+    {
+      "id": "evo2-tissue-identification",
+      "title": "Evo2 completed tissue-identification evaluation",
+      "snapshotPath": "provenance/evo2-tissue-20261102/entry.json",
+      "sha256": "6b1be888062990288244e8657a1462103e2aba687aca9152ee73496a0105766b"
+    },
+    {
+      "id": "evo2-tissue-identification-methods",
+      "title": "Evo2 extraction and independent evaluation checks",
+      "snapshotPath": "provenance/evo2-tissue-20261102/README.md",
+      "sha256": "7adc9a70eda713700bd91e5eaa23cc54cedb46ad6e5ffd5cb665a93e36b162cc"
+    },
+    {
+      "id": "evo2-tissue-identification-files",
+      "title": "Evo2 evaluation artifact checksums",
+      "snapshotPath": "provenance/evo2-tissue-20261102/SHA256SUMS.json",
+      "sha256": "a43ae2b3e1594858e8bd37f78165790cf5b63344394c6780f4faf9bc06180051"
+    }
+  ],
+  "evaluations": [
+    {
+      "modelId": "ceno_1b_1m",
+      "taskId": "opensplice",
+      "sourceId": "ceno-full-opensplice"
+    },
+    {
+      "modelId": "evo2_20b",
+      "taskId": "tissue_identification",
+      "sourceId": "evo2-tissue-identification"
+    }
+  ]
+}

source-snapshot/provenance/table-1.tex

@@ -0,0 +1,59 @@
+% Generated by build_benchmark_table.py from frozen source-linked estimates.
+\begin{table}[!ht]
+% Separate the table from the preceding Related works paragraph.
+\vspace{6pt}
+\centering\fontfamily{phv}\selectfont\fontsize{7.5}{8.5}\selectfont
+\hypersetup{hidelinks}
+\setlength{\tabcolsep}{1.5pt}\renewcommand{\arraystretch}{1.0}
+\definecolor{benchInk}{HTML}{26323D}
+\definecolor{benchRail}{HTML}{FFFFFF}\definecolor{benchRailHead}{HTML}{FFFFFF}
+\definecolor{benchPrior}{HTML}{A18338}
+\definecolor{benchSimulate}{HTML}{477FA5}
+\definecolor{benchContext}{HTML}{438879}
+\definecolor{benchCompose}{HTML}{8266A2}
+\definecolor{benchControl}{HTML}{B56E58}
+\color{benchInk}
+\newcommand{\benchhead}[2]{\textcolor{bench#1!85!black}{\mdseries\fontsize{6.8}{8.5}\selectfont #2}}
+\newcommand{\benchmetric}[1]{{\mdseries\fontsize{7.2}{8.7}\selectfont\shortstack{#1}}}
+\newcommand{\benchentry}[1]{\raisebox{-0.75pt}[0pt][0pt]{#1}}
+\begin{tabular}{@{}>{\columncolor{benchRail}[0pt][0pt]\raggedright\arraybackslash}p{\dimexpr.205\linewidth-1.5pt\relax}>{\columncolor{benchPrior!3}[0pt][0pt]\raggedleft\arraybackslash}p{\dimexpr.091\linewidth-3pt\relax}>{\columncolor{benchSimulate!3}[0pt][0pt]\raggedleft\arraybackslash}p{\dimexpr.088\linewidth-3pt\relax}>{\columncolor{benchContext!3}[0pt][1.5pt]\raggedleft\arraybackslash}p{\dimexpr.105\linewidth-3pt\relax}>{\columncolor{benchContext!3}[1.5pt][0pt]\raggedleft\arraybackslash}p{\dimexpr.085\linewidth-3pt\relax}>{\columncolor{benchCompose!3}[0pt][0pt]\raggedleft\arraybackslash}p{\dimexpr.11\linewidth-3pt\relax}>{\columncolor{benchControl!3}[0pt][1.5pt]\raggedleft\arraybackslash}p{\dimexpr.10\linewidth-3pt\relax}>{\columncolor{benchControl!3}[1.5pt][1.5pt]\raggedleft\arraybackslash}p{\dimexpr.111\linewidth-3pt\relax}>{\columncolor{benchControl!3}[1.5pt][0pt]\raggedleft\arraybackslash}p{\dimexpr.105\linewidth-3pt\relax}@{}}
+\multicolumn{1}{@{}>{\columncolor{benchRailHead}[0pt][0pt]}p{\dimexpr.205\linewidth-1.5pt\relax}}{\rule{0pt}{13pt}} & \multicolumn{1}{>{\columncolor{benchPrior!6}[0pt][0pt]}p{\dimexpr.091\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Prior}{\hyperref[fig:prior]{Fig.\,\ref*{fig:prior}A}}\par\vspace{4pt}\benchmetric{Prior\\loss\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont bits $\downarrow$}}\rule[-5pt]{0pt}{5pt}} & \multicolumn{1}{>{\columncolor{benchSimulate!6}[0pt][0pt]}p{\dimexpr.088\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Simulate}{\hyperref[fig:simulate]{Fig.\,\ref*{fig:simulate}A}}\par\vspace{4pt}\benchmetric{Splicing\\effects\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont $\rho\uparrow$}}\rule[-5pt]{0pt}{5pt}} & \multicolumn{1}{>{\columncolor{benchContext!6}[0pt][1.5pt]}p{\dimexpr.105\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Context}{\hyperref[fig:context]{Fig.\,\ref*{fig:context}B}}\par\vspace{4pt}\benchmetric{Cell\\contrast\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont $\rho\uparrow$}}\rule[-5pt]{0pt}{5pt}} & \multicolumn{1}{>{\columncolor{benchContext!6}[1.5pt][0pt]}p{\dimexpr.085\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Context}{\hyperref[fig:context-tissues]{Fig.\,\ref*{fig:context-tissues}B}}\par\vspace{4pt}\benchmetric{Tissue\\ID\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont \% $\uparrow$}}\rule[-5pt]{0pt}{5pt}} & \multicolumn{1}{>{\columncolor{benchCompose!6}[0pt][0pt]}p{\dimexpr.11\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Compose}{\hyperref[fig:compose]{Fig.\,\ref*{fig:compose}A}}\par\vspace{4pt}\benchmetric{Background\\interaction\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont $\rho\uparrow$}}\rule[-5pt]{0pt}{5pt}} & \multicolumn{1}{>{\columncolor{benchControl!6}[0pt][1.5pt]}p{\dimexpr.10\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Control}{\hyperref[fig:control]{Fig.\,\ref*{fig:control}A}}\par\vspace{4pt}\benchmetric{Single-edit\\rank\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont percentile $\downarrow$}}\rule[-5pt]{0pt}{5pt}} & \multicolumn{1}{>{\columncolor{benchControl!6}[1.5pt][1.5pt]}p{\dimexpr.111\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Control}{\hyperref[fig:control]{Fig.\,\ref*{fig:control}C}}\par\vspace{4pt}\benchmetric{Cell\\selectivity\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont $U\uparrow$}}\rule[-5pt]{0pt}{5pt}} & \multicolumn{1}{>{\columncolor{benchControl!6}[1.5pt][0pt]}p{\dimexpr.105\linewidth-3pt\relax}}{\centering\arraybackslash\rule{0pt}{13pt}\benchhead{Control}{\hyperref[fig:frontier-programming]{Fig.\,\ref*{fig:frontier-programming}A}}\par\vspace{4pt}\benchmetric{Multi-cell\\gain\\{\color{benchInk!80}\fontsize{6.5}{8}\selectfont log$_2$ $\uparrow$}}\rule[-5pt]{0pt}{5pt}}\\[3pt]
+\benchentry{Base composition} & \cellcolor{benchPrior!5.00}\benchentry{1.974\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{3pt}\mbox{}}\\
+\benchentry{Evo2 20B} & \cellcolor{benchPrior!31.47}\benchentry{1.388\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!6.86}\benchentry{0.367\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.39}\benchentry{0.041\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!6.92}\benchentry{0.050\hspace{4pt}\mbox{}} & \cellcolor{benchControl!7.27}\benchentry{43.5\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{-0.350\hspace{4pt}\mbox{}} & \cellcolor{benchControl!7.63}\benchentry{0.275\hspace{3pt}\mbox{}}\\
+\benchentry{CENO 1B} & \cellcolor{benchPrior!7.99}\benchentry{1.549\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.16}\benchentry{0.036\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.70}\benchentry{9.8\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.16}\benchentry{0.010\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.03}\benchentry{48.9\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.03}\benchentry{-0.318\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{-0.581\hspace{3pt}\mbox{}}\\
+\benchentry{GENERATORv2} & \cellcolor{benchPrior!5.00}\benchentry{1.711\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!7.70}\benchentry{0.388\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.78}\benchentry{0.066\hspace{4pt}\mbox{}} & \cellcolor{benchContext!6.45}\benchentry{17.6\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!6.29}\benchentry{0.046\hspace{4pt}\mbox{}} & \cellcolor{benchControl!6.57}\benchentry{44.7\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.32}\benchentry{-0.256\hspace{4pt}\mbox{}} & \cellcolor{benchControl!7.03}\benchentry{0.234\hspace{3pt}\mbox{}}\\
+\benchentry{DNABERT-2} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!5.16}\benchentry{0.198\hspace{4pt}\mbox{}} & \cellcolor{benchContext!6.28}\benchentry{0.076\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.06}\benchentry{5.9\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.39}\benchentry{0.020\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.13}\benchentry{48.6\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.13}\benchentry{-0.302\hspace{4pt}\mbox{}} & \cellcolor{benchControl!6.79}\benchentry{0.233\hspace{3pt}\mbox{}}\\
+\addlinespace[2.2pt]
+\benchentry{NTv3 100M pre} & \cellcolor{benchPrior!11.06}\benchentry{1.537\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!6.21}\benchentry{0.349\hspace{4pt}\mbox{}} & \cellcolor{benchContext!6.28}\benchentry{0.076\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.28}\benchentry{7.8\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.16}\benchentry{0.010\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.32}\benchentry{47.3\hspace{4pt}\mbox{}} & \cellcolor{benchControl!6.02}\benchentry{-0.207\hspace{4pt}\mbox{}} & \cellcolor{benchControl!8.54}\benchentry{0.339\hspace{3pt}\mbox{}}\\
+\benchentry{NTv3 100M post} & \cellcolor{benchPrior!15.15}\benchentry{1.492\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!12.53}\benchentry{0.502\hspace{4pt}\mbox{}} & \cellcolor{benchContext!11.81}\benchentry{0.244\hspace{4pt}\mbox{}} & \cellcolor{benchContext!16.34}\benchentry{51.0\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!9.05}\benchentry{0.139\hspace{4pt}\mbox{}} & \cellcolor{benchControl!14.14}\benchentry{30.3\hspace{4pt}\mbox{}} & \cellcolor{benchControl!11.38}\benchentry{0.211\hspace{4pt}\mbox{}} & \cellcolor{benchControl!6.18}\benchentry{0.144\hspace{3pt}\mbox{}}\\
+\benchentry{\hspace{6pt}\textcolor{black!65}{Native outputs}} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!5.00}\benchentry{0.168\hspace{4pt}\mbox{}} & \cellcolor{benchContext!8.06}\benchentry{0.203\hspace{4pt}\mbox{}} & \cellcolor{benchContext!7.76}\benchentry{27.5\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!10.40}\benchentry{0.152\hspace{4pt}\mbox{}} & \cellcolor{benchControl!11.22}\benchentry{32.0\hspace{4pt}\mbox{}} & \cellcolor{benchControl!10.35}\benchentry{0.200\hspace{4pt}\mbox{}} & \cellcolor{benchControl!9.13}\benchentry{0.344\rlap{$^{\ddagger}$}\hspace{3pt}\mbox{}}\\
+\addlinespace[1.8pt]
+\benchentry{NTv3 650M pre} & \cellcolor{benchPrior!22.22}\benchentry{1.447\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!5.03}\benchentry{0.187\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.00}\benchentry{0.010\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.28}\benchentry{7.8\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.03}\benchentry{0.009\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{50.4\hspace{4pt}\mbox{}} & \cellcolor{benchControl!7.47}\benchentry{-0.061\hspace{4pt}\mbox{}} & \cellcolor{benchControl!10.77}\benchentry{0.368\hspace{3pt}\mbox{}}\\
+\benchentry{NTv3 650M post} & \cellcolor{benchPrior!38.00}\benchentry{1.375\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!17.14}\benchentry{0.564\hspace{4pt}\mbox{}} & \cellcolor{benchContext!14.85}\benchentry{0.256\hspace{4pt}\mbox{}} & \cellcolor{benchContext!22.97}\benchentry{56.9\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!20.73}\benchentry{0.187\hspace{4pt}\mbox{}} & \cellcolor{benchControl!17.12}\benchentry{30.0\hspace{4pt}\mbox{}} & \cellcolor{benchControl!16.35}\benchentry{0.287\hspace{4pt}\mbox{}} & \cellcolor{benchControl!17.68}\benchentry{0.568\hspace{3pt}\mbox{}}\\
+\benchentry{\hspace{6pt}\textcolor{black!65}{Native outputs}} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!9.57}\benchentry{0.461\hspace{4pt}\mbox{}} & \cellcolor{benchContext!10.57}\benchentry{0.239\hspace{4pt}\mbox{}} & \cellcolor{benchContext!12.38}\benchentry{45.1\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!12.20}\benchentry{0.170\hspace{4pt}\mbox{}} & \cellcolor{benchControl!22.65}\benchentry{29.3\hspace{4pt}\mbox{}} & \cellcolor{benchControl!12.98}\benchentry{0.250\hspace{4pt}\mbox{}} & \cellcolor{benchControl!13.84}\benchentry{0.482\rlap{$^{\ddagger}$}\hspace{3pt}\mbox{}}\\
+\addlinespace[2.2pt]
+\benchentry{AlphaGenome} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!31.23}\benchentry{0.650\hspace{4pt}\mbox{}} & \cellcolor{benchContext!38.00}\benchentry{0.402\hspace{4pt}\mbox{}} & \cellcolor{benchContext!30.93}\benchentry{64.7\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!38.00}\benchentry{0.269\hspace{4pt}\mbox{}} & \cellcolor{benchControl!31.06}\benchentry{25.7\hspace{4pt}\mbox{}} & \cellcolor{benchControl!38.00}\benchentry{0.515\hspace{4pt}\mbox{}} & \cellcolor{benchControl!38.00}\benchentry{0.833\hspace{3pt}\mbox{}}\\
+\benchentry{\hspace{6pt}\textcolor{black!65}{Native outputs}} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!24.51}\benchentry{0.621\hspace{4pt}\mbox{}} & \cellcolor{benchContext!32.27}\benchentry{0.398\hspace{4pt}\mbox{}} & \cellcolor{benchContext!38.00}\benchentry{68.6\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!31.61}\benchentry{0.260\hspace{4pt}\mbox{}} & \cellcolor{benchControl!38.00}\benchentry{24.3\hspace{4pt}\mbox{}} & \cellcolor{benchControl!28.69}\benchentry{0.384\hspace{4pt}\mbox{}} & \cellcolor{benchControl!30.15}\benchentry{0.707\rlap{$^{\ddagger}$}\hspace{3pt}\mbox{}}\\
+\benchentry{Enformer} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchContext!20.69}\benchentry{0.282\hspace{4pt}\mbox{}} & \cellcolor{benchContext!9.17}\benchentry{31.4\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!15.44}\benchentry{0.181\hspace{4pt}\mbox{}} & \cellcolor{benchControl!12.39}\benchentry{31.5\hspace{4pt}\mbox{}} & \cellcolor{benchControl!21.41}\benchentry{0.290\hspace{4pt}\mbox{}} & \cellcolor{benchControl!24.76}\benchentry{0.700\rlap{$^{\ddagger}$}\hspace{3pt}\mbox{}}\\
+\addlinespace[2.2pt]
+\benchentry{$k$-mer ridge} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!5.65}\benchentry{0.318\hspace{4pt}\mbox{}} & \cellcolor{benchContext!9.00}\benchentry{0.211\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.28}\benchentry{7.8\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.76}\benchentry{0.032\hspace{4pt}\mbox{}} & \cellcolor{benchControl!6.07}\benchentry{45.0\hspace{4pt}\mbox{}} & \cellcolor{benchControl!6.73}\benchentry{-0.097\hspace{4pt}\mbox{}} & \cellcolor{benchControl!11.91}\benchentry{0.394\hspace{3pt}\mbox{}}\\
+\benchentry{Task-trained ResNet} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!11.09}\benchentry{0.488\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.03}\benchentry{0.020\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.00}\benchentry{2.0\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.00}\benchentry{-0.001\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.61}\benchentry{47.0\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.63}\benchentry{-0.215\hspace{4pt}\mbox{}} & \cellcolor{benchControl!9.82}\benchentry{0.347\hspace{3pt}\mbox{}}\\
+\benchentry{SpliceAI} & \cellcolor{benchPrior!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchSimulate!38.00}\benchentry{0.674\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchContext!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchCompose!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \cellcolor{benchControl!5.00}\benchentry{\textcolor{benchInk!27}{---}\hspace{3pt}\mbox{}}\\
+\addlinespace[3pt]
+\rowcolor{black!3}\cellcolor{white}\benchentry{\textcolor{black!65}{\fontsize{7.2}{8.5}\selectfont Experimental reference\rlap{$^{\dagger}$}}} & \benchentry{\textcolor{benchInk!27}{---}\hspace{4pt}\mbox{}} & \benchentry{0.809\hspace{4pt}\mbox{}} & \benchentry{0.594\hspace{4pt}\mbox{}} & \benchentry{99.0\hspace{4pt}\mbox{}} & \benchentry{0.421\hspace{4pt}\mbox{}} & \benchentry{11.8\hspace{4pt}\mbox{}} & \benchentry{0.962\hspace{4pt}\mbox{}} & \benchentry{1.602\hspace{3pt}\mbox{}}\\
+\end{tabular}
+\color{black}\rmfamily\setlength{\abovecaptionskip}{4pt}
+\caption{\textbf{The DRAFT benchmark.} Metrics from all experiments collected in one overview table. Referenced figures and their corresponding sections provide detailed information and explorations. Deeper shading indicates better performance within each column. %Cell selectivity reports absolute utility; multi-cell reports gain over intent-matched random (both log$_2$ units). References: uniform edit selection, no edit and random selection.
+%$k$-mers: 1--6 for multi-cell, 3--6 elsewhere. 
+Foundation-model results use embedding-fitted linear heads, except for prior loss (native sequence logits) and Enformer/indented rows (native outputs).
+% using models with an existing output head.
+%For functional tasks, Enformer and indented rows use the native functional outputs. 
+For multi-cell design, $\ddagger$ marks a linear head fitted to native outputs.
+$\dagger$: Experimental references compare donor or replicate subsets; background interaction uses HepG2--K562 transfer.
+%They are not ceilings: models may outperform these experimental measurements. 
+Measurement details are in Appendix~\hyperref[app:experimental-references]{\ref*{app:predictions-readouts}}. %---: not reported. %Definitions and uncertainty: Appendices~\ref{app:benchmark-measures}, \ref{app:gtex-tissue-contrast}; sources: \texttt{bend2-main-benchmark.csv}.
+}
+\label{tab:bend2-reference}
+% Balance the added space above without changing the page budget.
+\vspace{-6pt}
+\end{table}

task-mapping.md

@@ -0,0 +1,93 @@
+# BEND 2.0 native HF protocol map
+
+Eight scientific endpoints are preserved. Seven downstream endpoints have separate embeddings and native-output subleaderboards so a shared checkpoint never collapses two readouts into one score. The sequence-prior endpoint is independent of downstream readout and is published once per checkpoint.
+
+| Scientific task | Native task ID | Unit | Better |
+|---|---|---|---|
+| Prior loss | `prior_loss_bits_per_base` | bits/base | Lower |
+| Splicing effects | `opensplice_embeddings_spearman_rho` | ρ | Higher |
+| Splicing effects | `opensplice_native_outputs_spearman_rho` | ρ | Higher |
+| Cell contrast | `context_embeddings_spearman_rho` | ρ | Higher |
+| Cell contrast | `context_native_outputs_spearman_rho` | ρ | Higher |
+| Tissue identification | `tissue_identification_embeddings_accuracy_percent` | % | Higher |
+| Tissue identification | `tissue_identification_native_outputs_accuracy_percent` | % | Higher |
+| Background interaction | `background_embeddings_spearman_rho` | ρ | Higher |
+| Background interaction | `background_native_outputs_spearman_rho` | ρ | Higher |
+| Single-edit rank | `repair_percentile_embeddings_residual_midrank_percentile` | percentile | Lower |
+| Single-edit rank | `repair_percentile_native_outputs_residual_midrank_percentile` | percentile | Lower |
+| Cell selectivity | `cell_selective_embeddings_utility_log2` | log₂ utility | Higher |
+| Cell selectivity | `cell_selective_native_outputs_utility_log2` | log₂ utility | Higher |
+| Multi-cell gain | `atlas_profile_embeddings_gain_over_random_log2` | log₂ gain | Higher |
+| Multi-cell gain | `atlas_profile_native_outputs_gain_over_random_log2` | log₂ gain | Higher |
+
+## Prior loss
+
+Predict the hidden central base of 512 human reference-genome windows. Lower loss means better calibrated conditional probabilities.
+
+Metric: **Hidden-base conditional log loss**, in **bits/base**. 512 windows of 1,200 bp from separate 1-Mb genomic blocks; central base hidden.
+
+- 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.
+
+## Splicing effects
+
+Predict mutation-induced changes in alternative splicing; correlate within each held-out gene, then average genes equally.
+
+Metric: **Macro-gene Spearman correlation**, in **ρ**. 590,104 interventions in 466 genes; five folds hold out complete genes for fitted readouts.
+
+- 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.
+
+## Cell contrast
+
+Predict the HepG2-minus-K562 difference in a mutation’s reporter effect, then correlate within regulatory elements and average elements equally.
+
+Metric: **Macro-element Spearman correlation**, in **ρ**. 170,292 substitutions across 128 elements; complete elements and related allelic backgrounds remain together during fitting and evaluation.
+
+- The experimental reference has physical counts for 127 of 128 elements and uses repeat subsets; model values use the pooled posterior effects.
+
+## Tissue identification
+
+Match a predicted expression profile across held-out genes to the correct measured tissue among 51 candidates.
+
+Metric: **Top-1 tissue identification accuracy**, in **%**. 5,782 held-out genes and 51 represented tissues; 12,146 training and 1,350 validation genes.
+
+- 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.
+
+## Background interaction
+
+Predict how the same mutation’s effect differs across two parental genetic backgrounds, isolating background dependence.
+
+Metric: **Macro-element Spearman correlation**, in **ρ**. 68,026 paired mutations in 114 elements; paired backgrounds remain together.
+
+- 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.
+
+## Single-edit rank
+
+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.
+
+Metric: **Mean measured residual midrank percentile**, in **percentile**. 331 response targets in 110 elements; 578–597 eligible substitutions per element; direct reversion is excluded.
+
+- All candidate measurements are withheld until evaluation; predictions rank proximity to the model’s own predicted reference-edit effect.
+
+## Cell selectivity
+
+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|.
+
+Metric: **Mean cell-selective utility**, in **log₂ utility**. 167,958 substitutions, 280 parental backgrounds and 127 loci; both cells serve as targets; loci receive equal weight.
+
+- 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.
+
+## Multi-cell gain
+
+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.
+
+Metric: **Mean utility gain over intent-matched random selection**, in **log₂ gain**. 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.
+
+- 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.

validation.json

@@ -0,0 +1,24 @@
+{
+  "status": "local_structure_and_source_consistency_passed_not_hf_push_validation",
+  "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,
+  "lower_is_better_protocol_ids": [
+    "prior_loss_bits_per_base",
+    "repair_percentile_embeddings_residual_midrank_percentile",
+    "repair_percentile_native_outputs_residual_midrank_percentile"
+  ],
+  "result_key_uniqueness": true,
+  "all_104_source_cells_retained": true,
+  "missing_scores_omitted": true,
+  "negative_utilities_retained": true,
+  "source_precision_retained": true,
+  "combined_index_exported": false,
+  "verified_badge_claimed": false,
+  "datasets_or_model_repositories_mutated": false
+}