hukuang/plsdb-eex-esm2-context-embeddings

PLSDB short-protein ESM-2 context embeddings

ESM-2 (3B) embeddings for 2,427,482 unique plasmid proteins drawn from PLSDB, covering every 50–160 aa coding sequence used as a neighbourhood centre plus its ±3 flanking CDS. Built for entry-exclusion (EEx) gene discovery across the eight MPF conjugation classes, but the vectors are task-agnostic.

Each vector is the residue-mean of layer 36 over the full sequence, with no truncation, as float32[2560].

Contents

file	size	description
embedding_cache.sqlite	30.2 GB	one row per (sequence, signature)

Schema:

CREATE TABLE vectors (
  sha       TEXT,      -- sha256 of the uppercase AA sequence (see below)
  signature TEXT,      -- embedding protocol identifier
  vector    BLOB NOT NULL,  -- 10240 bytes = 2560 x float32, little-endian
  detail    TEXT NOT NULL,
  source    TEXT NOT NULL,
  PRIMARY KEY (sha, signature)
);

All 2,427,482 rows carry the single signature:

esm2_t36_3B_UR50D:layer36:residue_mean:full_sequence:v1

Key derivation

sha is the SHA-256 of the amino-acid string after: removing all whitespace, upper-casing, and dropping one trailing *. Only the 20 standard amino acids occur; any sequence with a non-standard residue was excluded, not substituted.

import hashlib, sqlite3, numpy as np

def key(seq):
    s = "".join(str(seq).split()).upper()
    return hashlib.sha256((s[:-1] if s.endswith("*") else s).encode()).hexdigest()

db = sqlite3.connect("file:embedding_cache.sqlite?mode=ro", uri=True)
row = db.execute("SELECT vector FROM vectors WHERE sha=?", (key(my_sequence),)).fetchone()
vec = np.frombuffer(row[0], dtype="<f4")   # shape (2560,)

How it was produced

• Model facebook/esm2_t36_3B_UR50D, layer 36, SDPA attention validated as full bidirectional before any inference.
• Pooling residue mean over the entire sequence. No chunking, no truncation, at any length. The longest input is 30,902 aa.
• Hardware NVIDIA A40, ~25 GPU-hours total, computed in four disjoint shards and merged.

Validation

Every figure below was measured on the delivered file, not asserted:

• PRAGMA integrity_check → ok
• 2,427,482 rows, exactly one signature value, no foreign-signature rows
• every vector: 2560-dim, float32, all values finite, non-zero norm

Two different shard statistics, deliberately kept apart

The work was split into four shards, and two overlap figures are reported about them. They are measured on different sets and must not be read as one statistic:

1. Assigned target sets — disjoint. Each sequence was assigned to exactly one shard by int(sha256[:8], 16) % 4. Measured on the per-shard input FASTA, with no database involved: union = sum = 2,427,482, every pairwise intersection 0, every sequence in the shard the rule assigns it to, nothing in the merged set missing from all shards and nothing in a shard absent from the merged set.

2. Delivered cache row sets — overlapping by exactly 20,000. Every shard was warm-started from the same 20,000-vector speed-test cache, so a shard's delivered SQLite contains its assigned targets plus that shared seed. The per-shard row counts are therefore 622,290 / 621,767 / 621,001 / 622,424, and each pairwise cache intersection is exactly the seed:

2,487,482 (sum of shard caches) − 3 × 20,000 (seed counted once) = 2,427,482

Merging asserted that the overlap equalled 20,000 at every step — not merely that it was small — and that sequences present in two shards carried bit-identical vectors, so no disagreement could be silently absorbed by an INSERT OR IGNORE.

• every transfer in the pipeline was content-verified: each file was hashed at source, copied, and re-hashed at the destination, and accepted only on agreement — size was never treated as sufficient

SHA-256 of embedding_cache.sqlite:

7ba15b69bd745777403179a3651826ca651484a030df957791146adf1f5458ac

Limitations

• Embeddings for sequences longer than 1022 aa are flagged as extended context. They are numerically valid; that is not the same as being biologically meaningful at those lengths, and no sensitivity analysis was run.
• Source CDS carry no functional annotation. PLSDB CDS headers in this build are pure coordinates, so nothing here encodes gene names or products.
• These are representations only. No classifier, score, or EEx label is included, and the presence of a sequence implies no functional claim.

Provenance and licensing

Sequences derive from PLSDB, which aggregates public INSDC plasmid records; consult PLSDB for its own citation and terms. The ESM-2 model is Meta AI's, released under MIT. This dataset contains derived numerical representations, not the source sequences — only their hashes.

If you use it, please also cite ESM-2 (Lin et al., Science 2023) and PLSDB.