Vocabulary interpretations of ESMC-6B SAE features

README.md

@@ -0,0 +1,145 @@
+---
+license: cc-by-4.0
+task_categories:
+  - feature-extraction
+tags:
+  - protein
+  - sparse-autoencoder
+  - interpretability
+  - annotation-vocabulary
+  - swiss-prot
+---
+
+# Vocabulary interpretations of ESMC-6B SAE features
+
+One row for every one of the 16,384 features of
+`biohub/ESMC-6B-sae-layer60-k64-codebook16384`, giving the protein annotation vocabulary term
+that best identifies what the feature detects, together with how well that identification holds on
+proteins the assignment never saw.
+
+This is the counterpart to `biohub/ESMC-SAE-Features`, produced without a language model. Where that
+release gives a free-text hypothesis per feature, this gives a controlled-vocabulary term plus a
+falsifiable prediction: which held-out proteins the feature should fire on, and how well it does.
+
+## What the numbers mean
+
+Every measurement here comes from one of four protein sets that never overlap, because splits are
+assigned to whole MMseqs2 clusters at 30% identity and 50% coverage:
+
+| Set | Role |
+|---|---|
+| Discovery, 26,725 train-split cluster representatives | fits the assignment |
+| Validation, 2,402 proteins in 1,582 clusters | chooses the ranking rule and support floor, nothing else |
+| Test, 2,257 proteins in 1,513 clusters | evaluated once against the frozen choice |
+| Biohub's released exemplars | never seen by any of the three; every cluster containing one is removed from all of them |
+
+A feature receives a term only if the term clears a threshold set by the largest association any
+feature achieves under a cluster-block permutation null over the full 16,384-feature family. That
+threshold is `z = 15.56`; a nominal Benjamini-Hochberg cut over the same tests would sit at 4.36.
+Among terms clearing it, the term with the highest AUROC wins, subject to a floor of 25 carriers.
+
+## Columns
+
+### `feature_interpretations.parquet`, 16,384 rows
+
+| Column | Meaning |
+|---|---|
+| `feature_id` | index into the codebook, 0 to 16,383 |
+| `status` | `assigned_and_evaluated`, `assigned_too_few_test_positives`, or `no_term_above_threshold` |
+| `term_id` | canonical term identity, `aspect:source_database:stable_id`; **use this as the key** |
+| `stable_id`, `source_database` | the parts `term_id` is composed from |
+| `term` | integer token, valid only for the vocabulary build recorded in `provenance.json` |
+| `n_total`, `assigned` | discovery population size, and whether a term cleared every threshold |
+| `label`, `aspect` | human-readable term name and which ontology it came from |
+| `z_score`, `auroc`, `n_carriers` | discovery-set association strength, effect size, and support |
+| `term_by_z`, `label_by_z`, `auroc_by_z` | the alternative ranking, for comparison |
+| `held_out_auroc` | AUROC on the cluster-disjoint test split |
+| `held_out_ap`, `held_out_ap_lift`, `held_out_prevalence` | average precision, and its lift over the base rate |
+| `held_out_recall_at_precision_50` | null for two different reasons: 1,828 evaluated features never reach precision 0.5, and 11,914 were not evaluated at all |
+| `held_out_n_positive` | test-split carriers; null where the feature was not evaluated |
+| `concordance_hits`, `concordance_drawn`, `concordance_enrichment`, `concordance_q` | agreement with Biohub's own 20 top-activating proteins for that feature |
+| `biohub_label` | the released agent label, for side-by-side reading |
+
+**Key on `term_id`, not on `term`.** Integer tokens serialize one vocabulary version and shift
+whenever an aspect is added or removed, which has just happened: the released 2026_02 build drops
+Gene3D, so every token above the removed range moves.
+
+### `local_associations.parquet`
+
+Residue-level associations between features and annotated spans, tested within each protein against
+a hypergeometric null and combined across proteins by Cochran-Mantel-Haenszel.
+
+| Column | Meaning |
+|---|---|
+| `feature_id`, `local_term`, `local_term_id`, `local_stable_id` | the pair. **The local vocabulary is not the global one**: `local_term` and `term` are different integer spaces and the two files must never be joined on them |
+| `label`, `aspect` | the local term's name and which UniProt feature key it came from |
+| `z_score`, `enrichment`, `observed`, `expected` | inside-span activity against the within-protein expectation |
+| `n_strata` | contributing proteins |
+| `above_positional_null_max` | **the column that matters**, see below |
+| `positional_null_max_z` | the guardrail value that flag is taken against, 61.96 |
+
+The file ships 88,717 pairs above a loose `z = 4`. That cut is not a finding: reshuffling every span
+to a random position inside its own protein, preserving length and count, still puts 6.0% of pairs
+above it. Filter on `above_positional_null_max` for the 2,907 pairs over 1,985 features that exceed
+every value the positional null produced across 507,098 reshuffled pairs. That is an observed
+guardrail, the largest value the null happened to reach, not a calibrated error rate.
+
+## How well the interpretations hold
+
+| | |
+|---|---|
+| Features with a term | 8,226 of 16,384 |
+| Distinct terms used | 354 |
+| **Features actually evaluated held-out** | **4,470 of the 8,226** |
+| Median held-out AUROC | 0.889, cluster-bootstrap 95% interval 0.872 to 0.913 |
+| Held-out AUROC above 0.8 | 63.2% of the 4,470 evaluated, 34.3% of the 8,226 assigned |
+| Median precision lift over base rate | 27.5x |
+| Concordance with Biohub's exemplars | 29.2x their Swiss-Prot base rate, 58.2% significant at q < 0.01, over the 8,224 features that could be scored |
+
+**Read the denominators.** A feature is evaluated held-out only if its term has at least five
+carriers among the 2,257 test proteins, which 4,470 of the 8,226 assigned features do. The other
+3,756 are not failures; the test split was too small to score them, and their held-out columns are
+null. Concordance is over 8,224 features, two assigned features having no scoreable exemplar set.
+
+## Limitations, stated rather than buried
+
+- **This is concordance, not accuracy.** Biohub publishes no evaluation of description quality and
+  its own tutorial calls the descriptions "automatically generated hypotheses". There is no ground
+  truth here to be right about, and none is claimed.
+- **8,226 features, 354 terms.** Roughly 10 features share each term at the median. Those features
+  are not duplicates, they nominate largely different proteins with median pairwise Jaccard 0.072,
+  but the number of distinct nameable concepts is bounded by the discovery set, not by the codebook.
+- **A vocabulary term is not a description.** Where a feature encodes something the vocabulary
+  cannot name, the method returns the nearest available term or nothing. Nine of the ten strongest
+  assignments resolve
+  distinct parts of a bacterial signal peptide all receive the single keyword `Signal`, which 120
+  features carry in total.
+- **The test split is cluster-disjoint but not pristine.** It has been evaluated under several run
+  configurations. What is claimed is narrower and is what the protocol enforces: the ranking rule and
+  support floor were chosen on validation alone, and no configuration was selected on test
+  performance.
+- **The association ran against the previous vocabulary**, which still contains Gene3D. Re-running
+  with Gene3D withheld gives 8,020 assignments at held-out AUROC 0.884 against 8,226 at 0.889, so the
+  difference is small, but it is a difference.
+- **Provenance is partial.** `provenance.json` records digests of the analysis inputs, the analysis
+  modules, and the run's own outputs, computed over at most the first 1 GiB of any file. It does not
+  contain digests of the six files shipped here, and the module digests describe the code as it
+  stood at the run, not necessarily the current repository.
+- **The `biohub_label` column is redistributed from Biohub's released feature table** and is
+  governed by that release's terms, not by the license on this dataset. Drop the column if that
+  matters for your use.
+- **Activations are recomputed, not the released ones.** Fused attention and layer-norm kernels were
+  unavailable, so values differ numerically. The released exemplars sit at median percentile 0.99986
+  of our activation distribution for their feature; the same check against an unrelated codebook
+  gives 0.464.
+
+## Provenance
+
+`provenance.json` carries SHA-256 digests of every input, every analysis module, the resolved
+options, and every output file, plus the seed and permutation count. Analysis code and the report
+that describes the method are in the accompanying repository.
+
+## Citation
+
+If you use this table, please also cite Biohub's ESMC and SAE releases, whose feature table and
+exemplar proteins make the comparison in this work possible.

feature_interpretations.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:352aae2485f670cdc9171b9e75e74c9e6dd7a2da34e30aa172abcadede2c6917
+size 863757

interpretations_summary.json

@@ -0,0 +1,11 @@
+{
+  "n_features": 16384.0,
+  "n_assigned_and_evaluated": 4470.0,
+  "n_assigned": 8226.0,
+  "median_held_out_auroc": 0.8890832589444182,
+  "aspects_used": 8,
+  "n_local_pairs": 88717.0,
+  "n_local_features": 13698.0,
+  "n_local_pairs_above_null_max": 2907.0,
+  "n_local_features_above_null_max": 1985.0
+}

local_associations.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfd9099ddd518f5d9f389419c937ba6b1432048318dcd64142bb5b50d41dbd28
+size 3452833

provenance.json

@@ -0,0 +1,54 @@
+{
+  "inputs": {
+    "run_dir": "sha256:ea30d91adbc1367b1038e972c7781a623a69f410018d0d4e1af1ceeb99325a85:3765190",
+    "protein_table": "sha256:0afd5b471f54a63109b82e3175aac2869ba9774f500a66b669a40e9aa1aa4aa5:26612759",
+    "term_table": "sha256:8e2f64b311fb5293978218f6becd80686b53a35b766fe2af01a5176b58e4644c:2724821",
+    "pooled": "sha256:0942c2bba0b1488c3b9eaaa00d485cbf28cb79acbda8c5d7080d8b63b4e5262f:67108864"
+  },
+  "code": {
+    "__init__.py": "sha256:41e82cc9acca3659fed8bd618b8340236801e0d29bd56fa5691f5d0b3b10ba77:256",
+    "aspect_ablation.py": "sha256:85353fe7cb36f3fa8755b78eb76cd14bd190479c5f8b5d5f7cab8c6b99b668b0:4795",
+    "aspect_redundancy.py": "sha256:8176c82d766008316fd4281f1ca5c0c0c881ed3f5811d5b6011d0c669d6aef70:5831",
+    "association.py": "sha256:c0df2a682e454f6dda5367bc24d47a8f9a0f86c3ad53b7780d56358f6747e04f:22549",
+    "biohub.py": "sha256:09195f39cb7d68d099c55317752a3de57089e10add7d9fe6b79c16080f093ecc:19002",
+    "examples.py": "sha256:0714d8dc01e27cb1c2775ccb8aae052ddde69f2b9aa648db8f07a9801f27e382:5692",
+    "export_interpretations.py": "sha256:133c040aa02b485449e3e23728361c7a6c32e9d0994b42e214e123723144ab11:5469",
+    "extraction.py": "sha256:3e24a39c60b05319070b8c3699b61d652a66c85bf2a2c069dcf454f2dea3c651:14885",
+    "fetch_labels.py": "sha256:993b338d3939901236a533b62b4756c321e7d500967f2b1fa388f0542bf9f50b:3156",
+    "fidelity_check.py": "sha256:276248bfcfbe49e87b0bf57bc4d932fcfe770a89b721f71824b64b9552e3d39a:5348",
+    "figures.py": "sha256:99f8726e9d4f1baa5ea2f7fb05ef8a189e1fb6803fb2ddbf21f15697a3a5752e:15332",
+    "implied_terms.py": "sha256:5fa06348ee4969f1f9e35fc4a9b6ce3234e1b4518c401f387c11e5543fc28b0e:9514",
+    "local_spans.py": "sha256:6ae9fdc3b69e4526d9fc43448d8b13c116e71180d1c288a5a643fe0efc978b78:11308",
+    "merge_runs.py": "sha256:d48260a4f968387fe0d6b7db4b902c8761ad748248eb12f95ae678704cef2183:4969",
+    "pipeline.py": "sha256:4c157d9c942e8f8d86b51d38639be830e7b678d670a5f39041d8a70c5c3bf410:55382",
+    "recompute_concordance.py": "sha256:ef4e876e893b07aa7b4b0d234f4ac56b8f93d7495d55264081018030166cdd32:4161",
+    "report_assets.py": "sha256:7879a20de0ec95188b30bca57ace3ce65e99bb8b2aec3c9a66389fb806dbda0d:16456",
+    "secrets_env.py": "sha256:403f775f37cf661c5e644977853cc9fe7de95477dde0cfe70a4bd480123dd65d:1463",
+    "sweep.py": "sha256:cfca7531e116b310799d268491f0c340a074e83a4f842b5c83bd1e18088de503:6638",
+    "uncertainty.py": "sha256:03921f93aee1a41a25b04b083cbd6603f8a8e40bebb02063490c932e0caf372a:5513",
+    "wandb_log.py": "sha256:a66fe8b3da54322872407ace181aec7518c12d001a8fb56821b00d5e7f6f1eb6:2449"
+  },
+  "options": {
+    "pooling": "max",
+    "min_auroc": 0.6,
+    "min_support_requested": 50,
+    "min_support_selected": 25,
+    "rank_by_selected": "auroc",
+    "exemplar_exclusion": "cluster",
+    "representative_rule": "mmseqs",
+    "n_permutations": 4,
+    "seed": 20260826
+  },
+  "outputs": {
+    "aspect_value.parquet": "sha256:5bc6c076971429b4bcd5f37f2bd55c89f56967d658ff0c8a6000c11456423f93:3920",
+    "category_recovery.parquet": "sha256:a5f08274a5519f6419b8f7c7cdc24016212461926e46fc5a35c5a2e12ec51415:57151",
+    "concordance.parquet": "sha256:c0b20e0dd6571e6260247c98063d4b904212f5c933ebd7e0e218187679f4bc29:227917",
+    "feature_assignment.parquet": "sha256:633948290fedacc99cf9ac10b398528827780477cd614afe4ebc7c54de2f82e3:284370",
+    "fidelity_percentile.npy": "sha256:9e5b0714ed3a83eaedea358c1ce631a9e122fa128a270282e120e92f8dc5b6e7:2621064",
+    "global_auroc.npy": "sha256:0179e78e35e09dfb32cf8817d4c16c10d6f6eae79f9b95a6e3fc59f9f5b6aeba:1073741824",
+    "held_out_metrics.parquet": "sha256:e59aed2f9a21cf14a17f1fe4f44d7fc7e79a4da34a9242019a94afa94e08d638:168261",
+    "lexical_agreement.parquet": "sha256:d68e4beaf6c594dd8243f4dc1a65e18cae64bbb5bf03f80d54bf5abf36afacdc:230056",
+    "null_z.npy": "sha256:4415a77208720aeb3660f61f87a0180ac963230b43f3f6eebab9cdca4b946590:262272",
+    "rule_selection.parquet": "sha256:1e713b608419ebc76c83a69fce556069da2ededef0984ba0e4bfa94aa66112a7:3763"
+  }
+}