ProtSent-V2 ESM-2 150M
Contrastively fine-tuned ESM-2 150M producing fixed-length protein embeddings where biological similarity maps to embedding proximity. Intended for retrieval, clustering, and nearest-neighbour transfer.
Retrained on a corpus decontaminated against the benchmark test sets. Predecessor: oriel9p/protsent-esm2-150M. Other scale: GrimSqueaker/ProtSent-V2-35M.
Usage
from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim
model = SentenceTransformer("GrimSqueaker/ProtSent-V2-150M")
emb = model.encode([
"MKTLLLTLVVVTIVCLDLGYT",
"MKTLLLTLVVVTIVCLDLGYN",
"AGWYRSPQEGLKPVDTFKDIV",
])
print(cos_sim(emb[0], emb[1:]))
Embeddings are mean-pooled over the final layer, dimension 640. Matryoshka heads at 64/128/256 are available by truncating the embedding.
Training data
Three sources, all decontaminated (see below). No DMS/ProteinGym component — unlike the V1 release, which included ProteinGym DMS pairs under a CoSENT loss.
| source | pairs used |
|---|---|
| Pfam families | 284,683 |
| AlphaFold DB (Foldseek clusters) | 8,612,331 |
| STRING-DB v12 PPI | 15,000,000 |
| total | 23,897,014 |
Pfam and AFDB pairs are sampled within clusters; STRING is a fixed 15M-pair subsample (seed 42) of the filtered pair table.
Decontamination
Every source was searched against the benchmark test sequences with MMseqs2 easy-search
(corpus as query, 40% identity, 80% coverage, --cov-mode 1) and matching sequences
removed before training.
| corpus | rows before | rows after | removed |
|---|---|---|---|
| Pfam | 28,530,684 | 27,929,772 | 2.11% |
| AlphaFold DB | 135,404,259 | 126,301,607 | 6.72% |
| STRING | 76,070,154 | 71,891,417 | 5.49% |
Filter targets: biomap-research/fold_prediction (remote homology) and
Synthyra/bernett_gold_ppi (PPI) test splits. The result was verified by semi-joining each
training file against the removal lists: zero flagged sequences remained.
SCOPe-40 was not a filter target — it has no train/test split, so filtering against it would remove nearly all domain sequences from the corpus.
Training configuration
| setting | value |
|---|---|
| backbone | ESM-2 150M (640 hidden, 30 layers) |
| loss | CachedMultipleNegativesRankingLoss |
| contrastive batch | 1024 per device |
| gather across devices | off |
| synthetic hard negatives | off |
| multi-dataset sampler | proportional |
| Matryoshka dims | 64 / 128 / 256 |
| max sequence length | 512 |
| optimiser | AdamW, LR 2e-4, cosine_with_min_lr |
| precision / attention | bf16, flash-attention-2 |
| hardware | 6x NVIDIA B300 |
| steps | 3,890 (one epoch) |
| gradient-cache mini-batch | 64 |
| warmup | 300 steps |
Training code: github.com/oriel9p/ProtSent,
train_esm2_150m.sh.
Results
SCOPe-40 structural retrieval, test split, self excluded, no-hit queries counted as failures. Restricted to the 1,693 of 2,207 queries that have a non-self same-family protein in the gallery.
| method | R@1 | R@10 | MAP |
|---|---|---|---|
| ESM-2 150M | 0.5535 | 0.7702 | 0.4236 |
MMseqs2 (-s 7.5) | 0.6556 | 0.7401 | 0.4098 |
HMMER (phmmer, -E 10, default filters) | 0.6970 | 0.7809 | 0.4747 |
| HMMER (phmmer, filters off — strongest) | 0.7525 | 0.8978 | 0.6067 |
| ProtSent-V1 150M | 0.6615 | 0.8943 | 0.6431 |
| ProtSent-V2 150M | 0.7431 | 0.9368 | 0.7042 |
Paired bootstrap over queries (10,000 resamples): V2 − V1 R@1 +0.0809 [+0.0602, +0.1022]; V2 − MMseqs2 R@1 +0.0868 [+0.0620, +0.1116], both excluding zero.
Against a maximally sensitive profile search, top-1 is not ours. With filters off, phmmer reaches R@1 0.7525 against this model's 0.7431. The embedding advantage is in ranking depth and MAP (R@10 0.9368 vs 0.8978, MAP 0.7042 vs 0.6067), at one forward pass per sequence and with indexable sub-linear search, rather than an all-vs-all profile comparison.
Remote homology (the task the corpus was filtered against), test split:
| model | 3-NN accuracy | linear-probe accuracy |
|---|---|---|
| ESM-2 150M | 0.5194 | 0.7500 |
| ProtSent-V1 150M | 0.7047 | 0.7401 |
| ProtSent-V2 150M | 0.6612 | 0.7503 |
Under 3-NN this model scores below V1 here. V1 was trained on a corpus containing sequences at ≥40% identity to this test set; removing them removed that advantage. Under a linear probe the ordering reverses.
Limitations
- Under a trained linear probe on the final layer, this model is roughly neutral to slightly worse than the stock ESM-2 backbone across a 23-task suite. The advantage is in nearest-neighbour geometry, not in information a trained readout can extract.
- The final layer is not the best pooling layer for property prediction. In a layer sweep on remote homology, an intermediate layer (~2/3 depth) scored higher for every model tested, including the stock backbone.
- V2 differs from V1 in more than decontamination: no hard negatives, proportional sampling, no DMS source, larger effective batch. It is not a controlled ablation of filtering alone.
- Only the remote-homology and PPI test sets were decontamination targets; other benchmark test sets were not filtered against.