INDEPENDENT LIFE SCIENCES BENCHMARKS
Bio-AI Model Leaderboard
Compare the scientific accuracy, structural resolution, inference latency, and licensing conditions of state-of-the-art biological foundation models across protein folding, de novo design, small-molecule docking, and single-cell genomics.
| Rank | Model & Creator | Task Domain | Primary Accuracy Benchmark | Secondary Metric | Compute Footprint | License | BioAtlas Profile |
|---|---|---|---|---|---|---|---|
| #1 | AlphaFold 3★ SOTA OverallGoogle DeepMind & Isomorphic Labs · Nature (2024) | Protein Structure | 84.2%CASP15 / lDDT-PLI | 1.42 ÅAll-Atom RMSD | Cloud API / A100 GPU | Non-Commercial / Server | View Specs → |
| #2 | ESM-3Top GenerativeEvolutionaryScale · Science (2024) | De Novo Design | 98.4%Simultaneous GFP Design | 1.18Sequence-Structure Perplexity | 8x H100 (98B) / 1x A100 (1.4B) | Open Weights (1.4B) / Commercial | View Specs → |
| #3 | RoseTTAFold All-AtomBaker Lab (UW IPD) · Science (2024) | Protein Structure | 76.8%Protein-Ligand RMSD < 2Å | 0.89Complex TM-score | 1x RTX 4090 / A100 | Open Source (Academic Free) | View Specs → |
| #4 | RFdiffusionStandard in De NovoBaker Lab (UW IPD) · Nature (2023) | De Novo Design | 88.5%De Novo Binder Success | 0.94Designability TM-score | 1x GPU (>= 16GB VRAM) | Open Source (BSD-3) | View Specs → |
| #5 | DiffDockFastest DockingMIT CSAIL · ICLR (2023) | Small-Molecule Docking | 38.2%PDBBind Top-1 RMSD < 2Å | 0.8s / ligandInference Speed | 1x GPU or CPU | Open Source (MIT) | View Specs → |
| #6 | ChromaGenerate Biomedicines · Nature (2023) | De Novo Design | 92.1%Substructure Conformity | 0.84Solubility Score | Cloud Enterprise | Proprietary | View Specs → |
| #7 | scGPTSingle-Cell LeaderWang Lab / Stanford · Nature Methods (2024) | Single-Cell / Genomics | 89.6%Cell Type Annotation Macro-F1 | 0.81Perturbation Prediction Pearson r | 1x A100 (40GB) | Open Source (MIT) | View Specs → |
| #8 | GeneformerBroad Institute · Nature (2023) | Single-Cell / Genomics | 87.4%Dosage Sensitivity F1 | 0.78Chromatin Accessibility Correlation | 1x V100 / A100 | Open Source (Apache 2.0) | View Specs → |
| #9 | ESMFoldMeta AI · Science (2023) | Protein Structure | 60x faster than AF2Fast Fold Speed | 81.2%lDDT Accuracy | 1x GPU | Open Source (Apache 2.0) | View Specs → |
Methodology & Benchmark Integrity
Metrics are curated from published, peer-reviewed literature (Nature, Science, Cell, ICLR) and blind community evaluations (CASP15, CAMEO, PDBBind). All evaluations reflect standardized test sets without overlapping training data.
Want your model benchmarked or updated? Verified lab PIs and company representatives can submit benchmark evaluation artifacts →