kevinlu4588/protein-esmc6b-embeddings

Protein ESMC 6B residue embeddings

All supported sequences have published embeddings; unsupported sequences remain explicit.

The complete sequence catalog contains 1,688,547 unique sequences and 396,182,273 residues. The current release v0007 records 1,686,441 encoded sequences, 1,686,441 published embeddings, 0 encoded but not yet published, 0 awaiting encoding, and 2,106 unsupported noncanonical strings. Counts reconcile over the full catalog, not a 600,000-cluster subset.

This snapshot describes embedding commit 24d862dc948fc4d00f8aa68c401c077c8ab96c13. Later shards may appear additively and are not silently included in these counts. Older catalogs and shards remain available unchanged. No queued upload counts as published. The manifest contains every included shard/index checksum and size.

Scientific alias metadata: complete. Alias-specific per-residue metadata: complete (2,269,488 alias rows). Read the completion provenance in metadata/aliases/v1/ and the exact scientific_metadata.residue_leaves[*].provenance_path listed in the release manifest. Where an additive v2 metadata correction exists, the manifest selects that complete version; older files remain unchanged. These are different tables: full alias metadata retains all original scientific fields as JSON plus searchable identity columns; residue tables retain source-specific confidence, masks and mappings without borrowing geometry between aliases. Missing arrays are explicit, including historical registry-only metadata. Completion here describes export of available fields, not coordinate completeness or new confidence assessments.

Chain and assembly identity

The full-chain embeddings are per-sequence features, not complex-conditioned features. Multiple source aliases or chain copies may reference one sequence_sha256; reuse of that embedding does not merge their identities. Original scientific metadata retains available PDB entry IDs, entity IDs, label_asym_id, auth_asym_id, residue mappings, and any already-present assembly, operator, or chain-instance fields. These fields may be nested in original_scientific_metadata_json; the searchable columns are not an exhaustive field list. Missing source fields remain absent or null, not inferred. No biological assembly, stoichiometry, or complex membership is inferred from a chain record. A future complex manifest can reference the current sequence SHAs while preserving separate chain instances and source-specific residue mappings. No new assembly acquisition or model training is part of this release.

Full encoder/tokenizer/config hashes and numerical policy are in provenance/encoder-v1.json; no model weights are included. Source-specific licensing is in ATTRIBUTION.md.

All upstream raw feature producers have completed durable delivery for the 1,686,441 supported sequences. This storage status is separate from the Hugging Face publication counts above.

Representation and sequence indexing

BF16 SafeTensors files contain embeddings of shape [sum(length), 2560]. Each Parquet index contains exact full amino-acid strings, sequence SHA-256, zero-based residue offsets and exclusive ends, source/tensor checksums, and encoder and eligibility identities. No chain is split between shards; the target is 2 GiB of raw tensor bytes per shard. The encoder is the final state after 80 layers in ESMC 6B, provided by biohub/ESMFold2 at revision 69869f737beffec5294845ede23db5fc0b4f509e. CLS and EOS are removed after encoding; no residue crops, padding or substitutions are introduced in publication. The exact BF16 content is preserved, not converted to FP16/FP32. No model weights are included.

Load the complete sequence catalog

from datasets import load_dataset
catalog = load_dataset('kevinlu4588/protein-esmc6b-embeddings',
                       'sequences', split='train')
print(catalog[0]['sequence'], catalog[0]['embedding_status'])

train is a file-loading split name, not a benchmark-isolated training split. For reproducibility, pass an immutable Hugging Face commit to revision=.

Load one published embedding

from huggingface_hub import HfApi, hf_hub_download
from safetensors import safe_open
import pyarrow.parquet as pq
repo = 'kevinlu4588/protein-esmc6b-embeddings'
api = HfApi()
revision = api.dataset_info(repo).sha  # save this immutable commit
files = api.list_repo_files(repo, repo_type='dataset', revision=revision)
indexes = sorted(p for p in files if p.startswith('embeddings/') and p.endswith('.parquet'))
if not indexes:
    raise RuntimeError('No embedding shards are published yet at this revision')
index = hf_hub_download(repo, indexes[0], repo_type='dataset', revision=revision)
row = pq.read_table(index).slice(0, 1).to_pylist()[0]
path = hf_hub_download(repo, row['shard_path'], repo_type='dataset', revision=revision)
with safe_open(path, framework='pt', device='cpu') as f:
    embedding = f.get_slice(row['tensor_key'])[row['residue_offset']:row['residue_end']]
assert tuple(embedding.shape) == (len(row['sequence']), 2560)

Scientific scope and limitations

The final master is untruncated. The 600,000-cluster selection is optional metadata, not an export cap. Identical complete sequence strings share one embedding, but source/structure aliases retain separate confidence, masks, residue mappings and versions; one alias's geometry is never substituted for another's. Fragments retain their original model bounds; missing tails are not assigned invented coordinates. Sequence eligibility does not establish coordinate completeness, finite backbone frames, training admission or biological independence.

Benchmark reservations cover 194 CAMEO-panel sequences plus 84 retrieved unique CASP14 sequences (including 51 evaluation inputs), totaling 278 benchmark sequences, and 474 additional reserved sequences. Screening used MMseqs2 release 18, identity >=0.30 and shorter-sequence coverage >=0.80, plus retained domain, strong-match and identifier reservations. This is not a claim that every structural cluster member was individually screened. Additional historical CAMEO coverage is unresolved; no universal pretraining-freshness or held-out-independence claim is made. The initial source cache contained 977 later-reserved and 19 unknown keys, all excluded from this publication's embedding claims. Upstream inherited 98/1/1 partitions are not benchmark isolation.

See ATTRIBUTION.md for source-specific licensing and citations. Publication identifier: protein-esmc6b-publication-v1.

Load source-specific metadata

from datasets import load_dataset
aliases = load_dataset('parquet', data_files={
    'train': 'hf://datasets/kevinlu4588/protein-esmc6b-embeddings/metadata/aliases/v1/*.parquet'
}, split='train', streaming=True)
# Join alias_key to residue tables, and sequence_sha256 to the sequence catalog.
# Do not join residue masks using sequence_sha256 alone when aliases differ.

Read the release manifest before loading metadata; absent completion provenance means that metadata is not yet published. Pin the same immutable repository commit for all files when reproducing a result.