BioAtlas finds candidate model families, pins exact versions, exposes benchmark claims, freezes one evaluation dataset and metric contract, dispatches only cleared compute adapters, and keeps uncertainty visible until the evidence supports a verdict.
Version selection is task-specific: for example, Chai-1 is kept for complex prediction rather than substituting Chai-2 design claims.
All-atom biomolecular complex prediction
Open complex structure + affinity-capable lineage
Multimodal biomolecular complex prediction
Open AF3-class biomolecular complex prediction
Top-performing CASP14 system; consult the primary paper for target-level metrics.
peer-reviewed · replication multiple-independent-usesCASP performance does not establish equal accuracy for every target class or drug-relevant complex.AlphaFold 2 and AlphaFold 3 require separate evaluation contexts.Open evaluation reports structure prediction and later affinity capabilities.
developer-reported · replication partialBoltz-1 structure claims and Boltz-2 affinity claims should be separated by version.Affinity performance depends strongly on target family and split design.Developer-reported AF3-class complex-prediction performance.
developer-reported · replication partialPreprint and developer-reported comparisons require independent reproduction.Benchmark protocol and entity coverage determine comparability.No normalized model-specific claim is currently sufficient for this task. BioAtlas keeps this as an evidence gap.
Large protein-ligand interaction resource with similarity-aware evaluation splits for proteins, ligands, pockets and interactions.
Pose, interaction and task-specific structure metricsPhysical and chemical validity checks for generated or docked protein-ligand poses.
Geometry, stereochemistry, clashes and interaction-validity checksEstimate first if desired. Submission uses the existing BioAtlas compute gateway and will fail closed when clearance, immutable artifacts, checkpoint identity or dataset rights are missing.
Define the scientific context, inspect versions/evidence, then estimate or run the same governed evaluation across eligible adapters.
No execution adapter
g6.2xlarge · nominal 18 min
g6.2xlarge · nominal 20 min
No execution adapter
Only enter or import metrics generated under the frozen protocol above. Blank fields deliberately produce “Insufficient evidence.”
| Model | Pose successhigher · threshold 70% | Physical validityhigher · threshold 90% | Failure ratelower · threshold 15% | Reproducibilityhigher · threshold 80% | Unresolved uncertaintylower · threshold 25% | Evidence qualityhigher · threshold 70% | Failure cases |
|---|---|---|---|---|---|---|---|
| AlphaFold 3no-governed-adapter | % | % | % | % | % | % | |
| Boltz-2clearance-required | % | % | % | % | % | % | |
| Chai-1clearance-required | % | % | % | % | % | % | |
| Protenixno-governed-adapter | % | % | % | % | % | % |
Required normalized metrics are incomplete · no-governed-adapter.
AlphaFold 3 · 2024 · PLINDER · similarity-controlled protein–ligand evaluationRequired normalized metrics are incomplete · clearance-required.
Boltz-2 · 2025 · PLINDER · similarity-controlled protein–ligand evaluationRequired normalized metrics are incomplete · clearance-required.
Chai-1 · 2024 · PLINDER · similarity-controlled protein–ligand evaluationRequired normalized metrics are incomplete · no-governed-adapter.
Version history not yet curated · PLINDER · similarity-controlled protein–ligand evaluation