Synthyra/ESMC-6B-SAE-Annotation-Vocabulary-Features

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.

The Hub exposes separate global and local configurations. global is the default; each configuration contains one train split backed by the correspondingly named Parquet table below.

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 the largest feature-wise best-term association observed in four cluster-block permutation draws over the full 16,384-feature family. That guardrail is z = 15.56; a nominal Benjamini-Hochberg cut over the same tests would sit at 4.36. Four permutation maxima cannot calibrate a low family-wise error rate, so z = 15.56 is a coarse, uncalibrated guardrail rather than a calibrated significance threshold. 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	V1-derived display composite, aspect:source_database:stable_id; it is not a canonical av::... identifier or a direct key into the 2026_02 term dictionary
stable_id, source_database	the parts term_id is composed from
term	raw V1 integer token for assigned rows; unassigned rows use the sentinel -1
n_total, assigned	held-out test population size, 2,257, the denominator of held_out_prevalence and non-null only for the 4,470 evaluated features; 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

Do not join term_id directly to the 2026_02 term dictionary. This archival file's display composite predates the current av::aspect::source_database::stable_accession identifier grammar. For assigned global rows, migrate the raw V1 term through the old_token key in the migration configuration of Synthyra/SwissProt-Annotation-Vocabulary at immutable revision 045af7cdd38e83462ebb10882b7506deff21e658. Only exact and replaced outcomes provide an automatic current new_term_id and new_token; every other status requires the handling described by that release. Exclude unassigned rows before the join because their term value is the -1 sentinel, not a biological token.

from datasets import load_dataset

features = load_dataset(
    "Synthyra/ESMC-6B-SAE-Annotation-Vocabulary-Features",
    "global",
    split="train",
    revision="03472f5ddce0cb5608f5c5b5e27b2f4bdc0a273c",
    token=False,
).to_pandas()
assigned = features.loc[features["assigned"].astype(bool)].copy()
assert assigned["term"].ge(0).all()

migration = load_dataset(
    "Synthyra/SwissProt-Annotation-Vocabulary",
    "migration",
    split="migration",
    revision="045af7cdd38e83462ebb10882b7506deff21e658",
    token=False,
).to_pandas()
migration = migration.rename(columns={"status": "migration_status"})

current = assigned.merge(
    migration[["old_token", "migration_status", "new_term_id", "new_token"]],
    left_on="term",
    right_on="old_token",
    how="left",
    validate="many_to_one",
    indicator=True,
)
assert current["_merge"].eq("both").all()
current["automatic_current_mapping"] = current["migration_status"].isin(
    ["exact", "replaced"]
)

At these pinned revisions, all 8,226 assigned feature rows join to the migration table: 5,482 are exact, 1,363 are replaced, and 1,381 are removed/unmapped. The first two groups, 6,845 rows, have an automatic current mapping.

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 V1-derived local pair. The local vocabulary is not the global one: do not join these identifiers or integers directly to either the global table or the 2026_02 release
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, and all nine 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. The table contains 1,378 Gene3D assignments across 42 Gene3D terms. Of those assigned features, 556 were evaluated on the held-out split and 822 had too few held-out positives to evaluate. 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. Gene3D is excluded from the released 2026_02 annotation vocabulary.
• The identifier columns use V1-derived grammar. None of the 8,226 assigned global term_id values or 88,717 local_term_id values uses the current canonical av::... grammar. Assigned global raw tokens migrate through migration.old_token as shown above; unassigned global rows use term = -1. The local tokens span 88,286 through 90,072, outside the migration table's 0-through-88,280 key range. No released automatic mapping exists for these description-specific local hypotheses, although a future manual semantic mapping remains possible.
• 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.