Atlas
Sign in

Benchmarks

45 benchmarks in 8 domains

Protein

8 benchmarks · zero-shot · 18 models
Protocol
Benchmarks: 8 of 8

Mean rank and coverage are recomputed over the benchmarks ticked here. None ticked means all.

Reset

18 models

 Ranked from 3 of 8 benchmarks (a third, at least 2)
# Model ↕ Mean rank ▼ Coverage ↕ Domain variant fitness, human PLM 0.5 (ProteinGym x Pfam) ↑ Spearman ρ (zero-shot, domain input) ↕ Domain variant fitness, human PLM 0.9 (ProteinGym x Pfam) ↑ Spearman ρ (zero-shot, domain input) ↕ Domain variant fitness, human PLM Clans-0.5 (ProteinGym x Pfam) ↑ Spearman ρ (zero-shot, domain input) ↕ Domain variant fitness, human PLM Clans-0.9 (ProteinGym x Pfam) ↑ Spearman ρ (zero-shot, domain input) ↕ GO function prediction (Swiss-Prot 2025H2) ↑ Fmax (molecular function, native head) ↕ Protein stability ΔΔG (Mega-scale) ↑ Spearman ρ (zero-shot) ↕ Protein subcellular localization (DeepLoc) ↑ Macro F1, 5 classes (native classifier) ↕ ProteinGym DMS zero-shot (Spearman) ↑ Spearman ρ (zero-shot) ↕
12 protein-ml/esm2-t12-35m-ur50d overlap 0.70 6/8 0.282 (8) * 0.282 (8) * 0.282 (10) * 0.168 (11) * · 0.278 (8) · 0.388 (9) *
11 protein-ml/amplify-350m overlap 0.67 6/8 0.189 (12) * 0.189 (12) * 0.330 (7) * 0.042 (13) * · 0.335 (4) · 0.416 (5) *
10 protein-ml/protgpt2 overlap 0.66 6/8 0.261 (10) * 0.261 (10) * 0.063 (13) * 0.277 (4) * · 0.345 (3) · 0.182 (11) *
9 protein-ml/prot-bert overlap 0.60 6/8 0.228 (11) * 0.228 (11) * 0.382 (2) * 0.145 (12) * · 0.308 (6) · 0.402 (6) *
8 protein-ml/esm2-t30-150m-ur50d overlap 0.59 6/8 0.263 (9) * 0.263 (9) * 0.289 (9) * 0.261 (6) * · 0.247 (10) · 0.423 (4) *
7 protein-ml/esm2-t6-8m-ur50d overlap 0.55 6/8 0.347 (5) * 0.347 (5) * 0.164 (11) * 0.281 (3) * · 0.256 (9) · 0.316 (10) *
6 protein-ml/progen2-small overlap 0.49 6/8 0.338 (6) * 0.338 (6) * 0.341 (5) * 0.283 (2) * · 0.116 (12) · 0.399 (8) *
5 protein-ml/esm2-t33-650m-ur50d overlap 0.44 6/8 0.310 (7) * 0.310 (7) * 0.306 (8) * 0.233 (8) * · 0.327 (5) · 0.425 (2) *
4 protein-ml/amplify-120m overlap 0.34 6/8 0.351 (4) * 0.351 (4) * 0.367 (4) * 0.189 (10) * · 0.296 (7) · 0.434 (1) *
3 protein-ml/esm2-t36-3b-ur50d overlap 0.31 6/8 0.369 (3) * 0.369 (3) * 0.338 (6) * 0.254 (7) * · 0.410 (1) · 0.402 (7) *
2 protein-ml/progen2-base overlap 0.18 6/8 0.476 (1) * 0.476 (1) * 0.463 (1) * 0.362 (1) * · 0.237 (11) · 0.424 (3) *
1 protein-ml/esm2-t48-15b-ur50d overlap 0.15 5/8 0.385 (2) * 0.385 (2) * 0.369 (3) * 0.263 (5) * · 0.402 (2) ··
Low coverage: below 3 of 8, or nothing to compare with. Listed, not ranked.
— protein-ml/human-protein-language-model-0-5 overlap 1.00 1/8 0.160 (13) * ·······
— protein-ml/human-protein-language-model-0-9 overlap 1.00 1/8 · 0.134 (13) * ······
— protein-ml/human-protein-language-model-clans-0-5 overlap 0.92 1/8 ·· 0.144 (12) * ·····
— protein-ml/human-protein-language-model-clans-0-9 overlap 0.67 1/8 ··· 0.223 (9) * ····
— protein-ml/esm-protein-localization overlap — 1/8 ······ 0.638 (1) * ·
— protein-ml/esm2-150m-protein-molecular-function — 1/8 ···· 0.868 (1) ···

Each cell is the model's value on that benchmark's board, with its rank there in brackets, as the benchmark's own leaderboard ranks it (its primary zero-shot metric, at its default version); hover for the 95% interval, n and source, click to open that board. A * marks a declared train/test overlap. Rows are ordered by mean normalised rank, (rank − 1) / (board size − 1) averaged over the benchmarks the model has a result on (a board of one entry does not count): ranks, never values, so no average mixes units. Filtering hides rows without re-ranking anyone.