Skip to main content
BIO·ATLASby Bayes Pharma.ai
Search
AccountStart evaluation
Task and benchmark claim graph

Separate the claim from the model.

A benchmark statement is meaningful only with its model version, task, dataset, split, metric, reporting source and caveats. BioAtlas groups claims only when those fields align instead of manufacturing a universal leaderboard.

Canonical claim records

Every curated claim has an inspectable page.

26 claims
26normalized benchmark claims
Protein structure prediction

AlphaFold 2 / 3

AlphaFold 3 · Google DeepMind
peer-reviewed
DatasetCASP14 / complex evaluations
→
SplitCASP14 blinded targets
→
Metric or taskStructure accuracy and confidence

Top-performing CASP14 system; consult the primary paper for target-level metrics.

Reported by: Independent community assessment and model developersReplication: multiple independent uses
Known caveats
  • CASP performance does not establish equal accuracy for every target class or drug-relevant complex.
  • AlphaFold 2 and AlphaFold 3 require separate evaluation contexts.
protein-structure|casp14-complex-evaluations|structure-accuracy-and-confidence|casp14-blinded-targetsPrimary source ↗
Protein structure prediction

RoseTTAFold / All-Atom

RoseTTAFold All-Atom · Institute for Protein Design, UW
peer-reviewed
DatasetCASP14 / complex modelling
→
SplitCASP14 and held-out structure targets
→
Metric or taskStructure prediction

Competitive three-track structure prediction reported in the primary publication.

Reported by: Model developers with community benchmark contextReplication: independent use documented
Known caveats
  • Results across RoseTTAFold generations are not interchangeable.
  • All-atom complex evaluation requires a task-specific benchmark.
protein-structure|casp14-complex-modelling|structure-prediction|casp14-and-held-out-structure-targetsPrimary source ↗
Biomolecular complex prediction

Chai-1 / Chai-2

Chai-2 · Chai Discovery
developer-reported
DatasetComplex and antibody-design evaluations
→
SplitPoseBusters and complex-evaluation sets
→
Metric or taskStructure / design performance

Developer-reported AF3-class complex-prediction performance.

Reported by: Model developersReplication: partial
Known caveats
  • Preprint and developer-reported comparisons require independent reproduction.
  • Benchmark protocol and entity coverage determine comparability.
complex-structure|complex-and-antibody-design-evaluations|structure-design-performance|posebusters-and-complex-evaluation-setsPrimary source ↗
Biomolecular complex prediction

Boltz-1 / Boltz-2

Boltz-2 · MIT (Barzilay & Jaakkola labs)
developer-reported
DatasetPoseBusters and affinity benchmarks
→
SplitPoseBusters and affinity evaluation sets
→
Metric or taskStructure and affinity

Open evaluation reports structure prediction and later affinity capabilities.

Reported by: Model developers and open community evaluationsReplication: partial
Known caveats
  • Boltz-1 structure claims and Boltz-2 affinity claims should be separated by version.
  • Affinity performance depends strongly on target family and split design.
complex-structure|posebusters-and-affinity-benchmarks|structure-and-affinity|posebusters-and-affinity-evaluation-setsPrimary source ↗
De novo protein or binder design

RFdiffusion

RFdiffusion2 · Institute for Protein Design, UW
experimental
DatasetExperimental binder validation
→
SplitProspective designed proteins and binders
→
Metric or taskDe novo protein design

The primary publication includes experimental validation of generated designs.

Reported by: Model developers with wet-lab validationReplication: experimental primary
Known caveats
  • Experimental success rates depend on design objective and filtering pipeline.
  • A generated backbone still requires sequence design and laboratory confirmation.
protein-design|experimental-binder-validation|de-novo-protein-design|prospective-designed-proteins-and-bindersPrimary source ↗
De novo protein or binder design

ESM3

ESM3 · EvolutionaryScale (now CZ Biohub)
peer-reviewed
DatasetGenerative protein evaluations
→
SplitGenerative protein evaluations and experimental fluorescent-protein example
→
Metric or taskSequence, structure and function

Reported multimodal generation includes an experimentally characterized designed protein.

Reported by: Model developersReplication: experimental primary
Known caveats
  • A demonstration protein is not a universal measure of design performance.
  • Hosted and open variants may differ in capability and access.
protein-design|generative-protein-evaluations|sequence-structure-and-function|generative-protein-evaluations-and-experimental-fluorescent-protein-examplePrimary source ↗
Protein–ligand pose prediction

DiffDock

Version history not yet curated · MIT (Barzilay & Jaakkola labs)
peer-reviewed
DatasetPDBBind
→
SplitPDBBind held-out complexes
→
Metric or taskTop-ranked docking pose

Peer-reviewed top-ranked docking-pose evaluation.

Reported by: Model developersReplication: independent benchmarks exist
Known caveats
  • Random and temporal splits can produce materially different estimates.
  • Pose accuracy is not binding-affinity accuracy.
binding-pose|pdbbind|top-ranked-docking-pose|pdbbind-held-out-complexesPrimary source ↗
Genomic sequence modelling

Evo / Evo 2

Evo 2 · Arc Institute + Stanford + NVIDIA
peer-reviewed
DatasetGenomic sequence evaluations
→
SplitHeld-out genomic sequence evaluations
→
Metric or taskDNA/RNA/protein generation

Peer-reviewed sequence-generation and prediction evaluations across biological scales.

Reported by: Model developersReplication: emerging
Known caveats
  • Generation quality does not establish biological function or safety.
  • Evo and Evo 2 require version-specific evaluation.
genomic-generation|genomic-sequence-evaluations|dna-rna-protein-generation|held-out-genomic-sequence-evaluationsPrimary source ↗
Cell perturbation prediction

State / Stack

Stack · Arc Institute
developer-reported
DatasetPerturbation prediction datasets
→
SplitPerturbation-prediction datasets
→
Metric or taskCell-state prediction

Open evaluation of cell-state prediction under perturbation.

Reported by: Model developersReplication: emerging
Known caveats
  • Cell line, tissue, dose and timepoint shifts can dominate performance.
  • In-silico perturbations require experimental validation.
cell-perturbation|perturbation-prediction-datasets|cell-state-prediction|perturbation-prediction-datasetsPrimary source ↗
Single-cell representation learning

scGPT

Version history not yet curated · University of Toronto (Bo Wang Lab)
peer-reviewed
DatasetSingle-cell downstream tasks
→
SplitSingle-cell downstream-task benchmarks
→
Metric or taskCell representation

Peer-reviewed evaluation across representation and downstream single-cell tasks.

Reported by: Model developersReplication: multiple independent uses
Known caveats
  • Performance varies by preprocessing, batch correction and downstream task.
  • Representation quality is not equivalent to causal perturbation prediction.
single-cell-representation|single-cell-downstream-tasks|cell-representation|single-cell-downstream-task-benchmarksPrimary source ↗
Antibody sequence modelling

IgLM / AntiBERTy

Version history not yet curated · Johns Hopkins (Gray Lab)
peer-reviewed
DatasetAntibody sequence evaluations
→
SplitAntibody sequence evaluation sets
→
Metric or taskAntibody language modelling

Peer-reviewed antibody language-modelling evaluation.

Reported by: Model developersReplication: emerging
Known caveats
  • Sequence plausibility does not guarantee affinity, specificity or developability.
  • Training-set lineage and germline distribution affect generalization.
antibody-language|antibody-sequence-evaluations|antibody-language-modelling|antibody-sequence-evaluation-setsPrimary source ↗
De novo protein or binder design

BoltzGen

BoltzGen · MIT / Boltz team
experimental
DatasetBinder-design evaluations
→
SplitSplit details not yet normalized
→
Metric or taskDesign success and structural quality

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
protein-design|binder-design-evaluations|design-success-and-structural-quality|unknown-splitPrimary source ↗
Antibody sequence modelling

RFantibody

RFantibody · Institute for Protein Design, UW
experimental
DatasetExperimental antibody design
→
SplitSplit details not yet normalized
→
Metric or taskEpitope-specific de novo design

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
antibody-language|experimental-antibody-design|epitope-specific-de-novo-design|unknown-splitPrimary source ↗
Integrated discovery platform

ESM-2

Version history not yet curated · Meta AI (FAIR)
peer-reviewed
DatasetProtein representation and structure evaluations
→
SplitSplit details not yet normalized
→
Metric or taskTransfer / language-model representation

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
platform-discovery|protein-representation-and-structure-evaluations|transfer-language-model-representation|unknown-splitPrimary source ↗
Integrated discovery platform

SaProt

Version history not yet curated · Westlake / Zhejiang collaborators
peer-reviewed
DatasetProtein understanding benchmarks
→
SplitSplit details not yet normalized
→
Metric or taskStructure-aware representation

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
platform-discovery|protein-understanding-benchmarks|structure-aware-representation|unknown-splitPrimary source ↗
Protein–ligand pose prediction

Uni-Mol / Uni-Mol2

Version history not yet curated · DP Technology / DeepModeling
developer-reported
DatasetMolecular representation/property benchmarks
→
SplitSplit details not yet normalized
→
Metric or task3D molecular representation

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
binding-pose|molecular-representation-property-benchmarks|3d-molecular-representation|unknown-splitPrimary source ↗
Genomic sequence modelling

DNABERT-2

Version history not yet curated · Multi-institution research team
developer-reported
DatasetGenomic downstream tasks
→
SplitSplit details not yet normalized
→
Metric or taskGeneral genomic representation

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
genomic-generation|genomic-downstream-tasks|general-genomic-representation|unknown-splitPrimary source ↗
Genomic sequence modelling

Caduceus

Version history not yet curated · Cornell / Tri Dao collaborators
developer-reported
DatasetLong-range genomic benchmarks
→
SplitSplit details not yet normalized
→
Metric or taskVariant / sequence prediction

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
genomic-generation|long-range-genomic-benchmarks|variant-sequence-prediction|unknown-splitPrimary source ↗
Genomic sequence modelling

Borzoi

Version history not yet curated · Calico / academic collaborators
peer-reviewed
DatasetFunctional-genomics and RNA-seq evaluations
→
SplitSplit details not yet normalized
→
Metric or taskSequence-to-RNA prediction

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
genomic-generation|functional-genomics-and-rna-seq-evaluations|sequence-to-rna-prediction|unknown-splitPrimary source ↗
Integrated discovery platform

RiNALMo

Version history not yet curated · University of Zagreb / A*STAR collaborators
peer-reviewed
DatasetRNA downstream and structure tasks
→
SplitSplit details not yet normalized
→
Metric or taskRNA representation / structure prediction

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
platform-discovery|rna-downstream-and-structure-tasks|rna-representation-structure-prediction|unknown-splitPrimary source ↗
Genomic sequence modelling

LucaOne

Version history not yet curated · BioMap / collaborators
peer-reviewed
DatasetDNA/RNA/protein downstream tasks
→
SplitSplit details not yet normalized
→
Metric or taskCross-domain biological representation

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
genomic-generation|dna-rna-protein-downstream-tasks|cross-domain-biological-representation|unknown-splitPrimary source ↗
Cell perturbation prediction

Universal Cell Embeddings (UCE)

Version history not yet curated · Stanford / collaborators
peer-reviewed
DatasetCross-dataset single-cell evaluations
→
SplitSplit details not yet normalized
→
Metric or taskUniversal cell representation

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
cell-perturbation|cross-dataset-single-cell-evaluations|universal-cell-representation|unknown-splitPrimary source ↗
Cell perturbation prediction

AIDO Cell 1.0

Version history not yet curated · GenBio AI
developer-reported
DatasetVirtual Cell Benchmark 1.0
→
SplitSplit details not yet normalized
→
Metric or task31 metrics across five task families

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
cell-perturbation|virtual-cell-benchmark-1-0|31-metrics-across-five-task-families|unknown-splitPrimary source ↗
Integrated discovery platform

AIDO.Tissue

Version history not yet curated · GenBio AI
developer-reported
DatasetSpatial transcriptomics downstream evaluations
→
SplitSplit details not yet normalized
→
Metric or taskCell/tissue representation and prediction

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
platform-discovery|spatial-transcriptomics-downstream-evaluations|cell-tissue-representation-and-prediction|unknown-splitPrimary source ↗
Integrated discovery platform

GenBio-PathFM

Version history not yet curated · GenBio AI
developer-reported
DatasetTHUNDER / HEST / PathoROB
→
SplitSplit details not yet normalized
→
Metric or taskHistopathology representation and downstream performance

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
platform-discovery|thunder-hest-pathorob|histopathology-representation-and-downstream-performance|unknown-splitPrimary source ↗
Cell perturbation prediction

Tahoe-x1

Version history not yet curated · Tahoe Therapeutics
developer-reported
DatasetFour disease-relevant single-cell evaluation groups
→
SplitSplit details not yet normalized
→
Metric or taskEssentiality, cancer hallmarks, cell type and perturbation response

A structured benchmark claim is recorded; consult the linked source for numeric values and protocol details.

Reported by: Source authorsReplication: unknown
Known caveats
  • Protocol, split and implementation details must match before comparing this claim with another result.
cell-perturbation|four-disease-relevant-single-cell-evaluation-groups|essentiality-cancer-hallmarks-cell-type-and-perturbation-response|unknown-splitPrimary source ↗