Atlas
Sign in

Benchmarks

45 benchmarks in 8 domains

Protein

6 benchmarks · probe · 17 models
Protocol
Benchmarks: 6 of 6

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

Reset

17 models

 Ranked from 2 of 6 benchmarks (a third, at least 2)
# Model ↕ Mean rank ▲ Coverage ↕ GO function prediction (Swiss-Prot 2025H2) ↑ Fmax (molecular function, probe) ↕ Protein membrane association (DeepLoc) ↑ Q2 accuracy (logistic probe) ↕ Protein stability ΔΔG (Mega-scale) ↑ Spearman ρ (ridge probe) ↕ Protein subcellular localization (DeepLoc) ↑ Accuracy (logistic probe) ↕ ProteinGym DMS zero-shot (Spearman) ↑ Spearman ρ (ridge probe) ↕ Secondary structure (Q3) ↑ Q3 accuracy (per-residue probe) ↕
1 protein-ml/prot-t5-xl-half-uniref50-enc overlap 0.14 6/6 0.654 (10) 0.861 (2) * 0.297 (1) 0.754 (3) * 0.648 (2) * 0.813 (1) *
1 protein-ml/prot-t5-xl-uniref50 overlap 0.14 6/6 0.654 (10) 0.861 (2) * 0.297 (1) 0.754 (3) * 0.648 (2) * 0.813 (1) *
3 protein-ml/esm-protein-localization overlap 0.32 6/6 0.691 (6) 0.860 (4) * 0.224 (4) 0.767 (2) * 0.545 (10) * 0.739 (11) *
4 protein-ml/esm2-t33-650m-ur50d overlap 0.34 6/6 0.619 (16) 0.849 (6) * 0.253 (3) 0.728 (6) * 0.619 (4) * 0.797 (4) *
5 protein-ml/esm2-t30-150m-ur50d overlap 0.41 6/6 0.652 (12) 0.855 (5) * 0.094 (13) 0.698 (7) * 0.604 (5) * 0.803 (3) *
6 protein-ml/esm2-35m-protein-cellular-component overlap 0.42 6/6 0.683 (8) 0.888 (1) * 0.166 (6) 0.801 (1) * 0.451 (17) * 0.678 (13) *
7 protein-ml/esm2-t36-3b-ur50d overlap 0.44 6/6 0.645 (14) 0.848 (8) * 0.126 (9) 0.741 (5) * 0.600 (7) * 0.779 (5) *
8 protein-ml/esm2-150m-protein-molecular-function overlap 0.45 6/6 0.844 (1) 0.849 (6) * 0.116 (10) 0.677 (8) * 0.482 (15) * 0.749 (9) *
9 protein-ml/esm2-t12-35m-ur50d overlap 0.48 6/6 0.665 (9) 0.844 (9) * 0.155 (8) 0.655 (12) * 0.599 (8) * 0.773 (6) *
10 protein-ml/progen2-base overlap 0.57 6/6 0.705 (5) 0.824 (14) * 0.111 (11) 0.614 (15) * 0.661 (1) * 0.659 (15) *
11 protein-ml/esm2-35m-protein-biological-process overlap 0.58 6/6 0.715 (2) 0.837 (10) * 0.102 (12) 0.675 (10) * 0.497 (14) * 0.674 (14) *
12 protein-ml/esm2-t6-8m-ur50d overlap 0.61 6/6 0.684 (7) 0.830 (11) * 0.191 (5) 0.613 (16) * 0.476 (16) * 0.741 (10) *
12 protein-ml/protgpt2 overlap 0.61 6/6 0.707 (4) 0.822 (15) * 0.160 (7) 0.677 (9) * 0.532 (13) * 0.616 (17) *
14 protein-ml/amplify-120m overlap 0.67 6/6 0.651 (13) 0.822 (15) * 0.083 (14) 0.637 (14) * 0.601 (6) * 0.751 (8) *
15 protein-ml/amplify-350m overlap 0.71 6/6 0.640 (15) 0.828 (13) * 0.032 (17) 0.656 (11) * 0.544 (11) * 0.764 (7) *
16 protein-ml/progen2-small overlap 0.75 6/6 0.711 (3) 0.811 (17) * 0.034 (16) 0.585 (17) * 0.580 (9) * 0.625 (16) *
17 protein-ml/prot-bert overlap 0.77 6/6 0.566 (17) 0.830 (11) * 0.058 (15) 0.651 (13) * 0.541 (12) * 0.722 (12) *
Reference floors: trivial baselines on the same items, not hub models, unranked.
— kmer-protein floor 0.479 0.723 0.090 0.407 0.326 ·
— onehot-protein floor ···· 0.490 0.622

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 probe 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.