ching-goodfire/MAPS-ClinVar-VKS-Inputs

ESM-C layer-80 pathogenicity harvest — slim input manifest

Two normalized tables that carry the full information content of the original 406.8 MB harvest_manifest.parquet in 9.8 MB, plus the ClinVar identifiers needed to trace any variant back to its source record.

Files

file	rows	contents
variants.parquet	200,913	one row per variant: identity, split, label, cluster, ClinVar ids
wt_sequences.parquet	17,231	one row per RefSeq protein: protein_id, sequence
verification_report.json	—	every integrity check below with its measured numbers
checksums.txt	—	SHA-256 of both parquet files

variants.parquet schema

• variant_id — <RefSeq accession>_<wt><pos1><mut>, e.g. NP_000005.3_D277N. Unique, and the join key to the harvested tensors.
• protein_id — RefSeq protein accession; join key to wt_sequences.parquet.
• pos1 — 1-based residue position of the substitution.
• wt_aa, mut_aa — single-letter wild-type and mutant residues.
• seq_len — length of the wild-type protein (computed from the sequence, so never null; agrees with the ClinVar table's seq_len on all 192,973 joinable rows).
• GeneSymbol, VariationID — ClinVar gene symbol and VariationID; null for the 7,940 rows absent from the labeled ClinVar table (see caveats).
• label — 1 = pathogenic / likely pathogenic, 0 = benign / likely benign.
• split — test, train_existing, or train_new.
• source, stars — provenance flag and the ClinVar review-status star rating.
• cluster — MMseqs2 sequence-identity cluster id; splits are cluster-disjoint.
• mut_seq_override — null for every row whose mutant sequence is exactly recoverable by the rule below. Non-null rows carry their stored mutant sequence verbatim. Measured: 200,913 of 200,913 rows reconstruct exactly, 0 carry an override.

Reconstructing a mutant sequence

Every variant is a single amino-acid substitution, which is why the mutant sequences do not need to be stored:

import pyarrow.parquet as pq

v = pq.read_table("variants.parquet").to_pandas()
seqs = dict(pq.read_table("wt_sequences.parquet").to_pandas().values)

def mut_seq(row):
    if row.mut_seq_override is not None:
        return row.mut_seq_override
    wt = seqs[row.protein_id]
    i = row.pos1 - 1
    return wt[:i] + row.mut_aa + wt[i + 1:]

This rule was verified byte-for-byte against the stored mut_seq column on all 200,913 rows before that column was dropped, and a 5,000-row random sample was rebuilt into the original 9-column schema and compared for exact equality on every column.

What these inputs were used for

The layer-80 difference tensors are described here for context only and are not included in this release. This dataset is the input manifest; the tensors are a separate, much larger artifact.

For anyone who does have them: variant_id is the join key to <variant_id>.pt under the harvest's diff/L80/ directory — 200,913 files, one per row, verified against this variants table in both directions (0 rows without a tensor, 0 tensors without a row).

Each file holds the ESM-C layer-80 difference field restricted to the 256 residues with the largest mutant-minus-wild-type change: a 256 x 2560 fp16 tensor, an idx vector naming which residue positions those 256 rows are, and Xmean, the mean over the difference field. Layer 80 means the residual stream at block 80 of ESM-C 6B.

Splits

Splits are disjoint at both the protein and the MMseqs2 cluster level, so a protein — or any homolog of it — appears in exactly one split.

split	rows	proteins	clusters
test	12,565	581	494
train_existing	49,719	1,922	1,603
train_new	138,629	16,540	10,484

Test↔train overlap is 0 proteins and 0 clusters, reproduced independently from these tables rather than trusted from the upstream report.

Caveats

1. The 7,940 null GeneSymbol/VariationID rows are a provenance artifact, not a data error. Every one of the 7,940 rows absent from clinvar_missense_labeled.parquet carries source = existing, and 0 of the 138,629 clinvar_new rows are unjoined. So that table is the new ClinVar pull, and these 7,940 are carried-over variants from the earlier ProteinGym-derived clinical set which the pull did not include. The rate is 12.75% of the 62,284 existing rows, spread evenly rather than concentrated by split: 12.23% of test and 12.88% of train_existing. An accession-version explanation was tested and ruled out — 0 of 7,940 match another version of the same accession, and among the 1,286 rows checkable at the substituted position the wild-type residue disagrees in 0. The affected rows are complete in every other column, including label, split, cluster, and the substitution itself; only the two ClinVar identifier columns are null.

   The accounting closes in both directions: 192,973 joined + 7,940 manifest-only = 200,913, and 10,769 − 7,940 = 2,829, exactly the 203,742 − 200,913 row-count gap. The two exclusive sets are independent, not two spellings of one set.

2. The harvested population is a mildly class-skewed subsample of the labeled ClinVar slice. The 10,769 labeled ClinVar variants with no manifest row are not a random subsample: 76.69% are 1-star against 68.05% in the harvested population, and 40.25% are pathogenic against 28.24% — a 1.43x enrichment in pathogenic among the dropped rows. The cause of that drop is still unexplained. Anyone computing prevalence, base rates, or absolute calibration from this dataset should account for it. It does not affect probe evaluation, which rests on the cluster-disjoint test split with its own known label balance of 6,095 pathogenic / 6,470 benign.

3. The ClinVar release string is unrecoverable. No release identifier survives in the harvest outputs. The release is identified only by a count fingerprint from assembly_report.json: 11,050,882 missense candidates over 2,518,181 missense VariationIDs, 4,526,725 variant_summary IDs, 937,493 labeled pathogenic/benign, 203,742 final rows (58,835 pathogenic / 144,907 benign) over 17,845 proteins and 15,610 genes.

4. No harvest script survives for the tensors. The diff/L80/ tensors cannot be regenerated from code in the source tree; they are treated as a fixed input. This manifest describes them but does not reproduce them.

Two smaller notes: 10 variants were dropped upstream for wild-type residue mismatches, and 2,595 of the 17,231 proteins here exceed 1,022 aa (covering 91,221 variants), which is relevant to how the harvest windowed long sequences. Protein lengths run 32–35,991 aa, median 490.

Provenance

Built from /mnt/data/artifacts/proteins/layer80harvest/data/harvest_manifest.parquet and clinvar_missense_labeled.parquet, cross-checked against split_report.json and assembly_report.json. Annotation probes and the layer-78 trees are out of scope; this is the pathogenicity side only.