PawanRamaMali/proteingym-fm-benchmark

Protein Foundation Model Benchmark Results

Zero-shot fitness prediction results for protein foundation models evaluated on the ProteinGym substitution benchmark (217 DMS assays, ~2.7M variants).

Companion data for the paper: "From Sequence Encoders to Multimodal Systems: A Critical Survey of Protein Foundation Models" (IEEE TCBB 2026).

Dataset configurations

The dataset viewer exposes two configurations, because the files carry two different schemas that must not be merged:

• per_assay_summary (default): one row per DMS assay, with assay, spearman_rho, n_variants, n_scored, and reason (for skipped assays). Each model is a separate split (esm2_650m, esm2_3b, esm1v_ensemble, esm3_1_4b, prott5_xl, saprot_650m_v3/v4/v6, plus a partial esm2_3b run). This is the headline data behind the paper's Table.
• per_variant_scores: one row per variant, with mutant, DMS_score (experimental), and model_score (model prediction). Currently available for ESM-2 3B (201 assays under results/esm2_3B/scored/).

from datasets import load_dataset
# per-assay summary for SaProt v6 (the corrected SaProt result)
d = load_dataset("PawanRamaMali/proteingym-fm-benchmark",
                 "per_assay_summary", split="saprot_650m_v6")
# per-variant predictions for ESM-2 3B
v = load_dataset("PawanRamaMali/proteingym-fm-benchmark",
                 "per_variant_scores", split="esm2_3b")

Headline results

Model	Assays scored	Mean Spearman ρ	Median Spearman ρ
ESM-2 650M	201	0.446	0.493
ESM-2 3B	201	0.440	0.475
ESM-1v ensemble (5 seeds)	201	0.441	0.477
ESM3 1.4B (open)	201	0.428	0.473
ProtT5-XL-U50 (embedding-distance baseline)	134	0.296	0.282
SaProt 650M v3 (raw 3Di; superseded, see v6)	201	0.277	0.258
SaProt 650M v4 (pLDDT-masked; superseded, see v6)	201	0.302	0.333
SaProt 650M v6 (domain-aware windowed pLDDT-masked) CORRECTED	201	0.432	0.477

A note on the SaProt versions

Three SaProt entries appear above because earlier versions had a subtle data-pipeline bug. For assays that probe a sub-fragment of a larger UniProt entry (notably Tsuboyama-2023), v3/v4 truncated the 3Di token string to the first N residues of the protein rather than the assayed window, feeding the model tokens from a completely different part of the protein.

v6 is the corrected number. It substring-aligns each assay's WT sequence to the full UniProt sequence before windowing the 3Di tokens and pLDDT. After the fix, SaProt sits at ~97% of ESM-2's mean ρ (0.432 vs 0.446) and slightly above on the median (0.477 vs 0.475), matching the published claim of competitive performance.

File layout

per_assay_results/
  esm2_650M.csv
  esm2_3B.csv
  esm2_3B_partial.csv     # partial early run, kept for provenance
  esm1v_ensemble.csv
  esm3_1.4B.csv
  prott5_xl.csv
  saprot_650M_v3.csv      # superseded
  saprot_650M_v4.csv      # superseded
  saprot_650M_v6.csv      # correct SaProt result
results/
  <model>/per_assay_spearman.csv    # raw per-assay tables (same schema as per_assay_results)
  esm2_3B/scored/<assay>.csv        # per-variant predictions (mutant, DMS_score, model_score)
README.md                 # this file

Per-assay CSV columns: assay, spearman_rho, n_variants, n_scored, reason (present for skipped assays). Per-variant CSV columns: mutant, DMS_score, model_score.

Methodology

Scoring (five of six models): masked-marginal log-likelihood. For mutation A123V:

1. Mask position 123 in the WT sequence.
2. Run the masked sequence through the model.
3. Score = log P(V | context) - log P(A | context).

Multi-mutants score by summing single-mutation scores. This is the protocol used for ESM-2 650M, ESM-2 3B, ESM-1v (5-seed ensemble), SaProt (all versions), and ESM3 1.4B.

ProtT5 is the exception: it was scored by an embedding-distance heuristic (negative L2 distance between wild-type and mutant T5 encoder embeddings at the mutated positions), not a masked-marginal likelihood. Its value should be read as an embedding-distance baseline and is not strictly comparable to the likelihood-based scores of the other five models.

Metric: Spearman ρ between predicted scores and experimental DMS fitness, per assay; mean / median reported across all scored assays.

Skipped assays: ProteinGym contains a handful of proteins longer than the context windows of some models (>1022 aa for ESM-class encoders). These appear in the CSVs with a reason field but no ρ.

Source code

The benchmark scripts live at https://github.com/PawanRamaMali/protein-fm-benchmark. The v6 SaProt pipeline is in scripts/run_saprot_v6_domain.py.

Hardware

NVIDIA RTX A6000 (48 GB). Total wall-clock across all models: roughly 35-40 days.

Citation

@article{mali2026protein,
  title={From Sequence Encoders to Multimodal Systems: A Critical Survey of Protein Foundation Models},
  author={Mali, Pawan Rama and Kumar, Abhinav and Bharti, Vandana},
  journal={IEEE Transactions on Computational Biology and Bioinformatics},
  year={2026}
}

License

MIT for benchmark code and results; underlying model weights are subject to their respective licenses.