Skip to main content
BIO·ATLASby Bayes Pharma.ai
Search
AccountStart evaluation
BioAtlas Compass

Begin with the scientific problem—not a model name.

Describe the modality, objective, available data, access requirements, deployment and evidence expectations. Compass returns a transparent candidate stack, typed handoffs, alternatives, missing evidence and the experiments needed before trusting the output.

Recommended stack

5 selected models

Recommendations use transparent domain, capability, input-contract, access, evidence, deployment and compute matches. Scores rank catalogue fit; they are not scientific validity or readiness scores.

2typed errors8warnings
01
Target biology

Owkin

Owkin
15 match signals

Define or prioritize the biological target and evidence base.

Acceptsdisease contexttarget hypothesisprotein sequencedna rna sequence
→
Producestarget hypothesismultimodal evidenceprogram recommendationcandidate molecules
  • Fits the small molecule modality.
  • Fits the end to end objective.
  • Matched capabilities: platform, discovery, pipeline, multimodal.
  • Accepts at least one available project data type directly.
Cautions
  • No normalized benchmark claim has been curated yet.
Evidence 1/7 documented · benchmark not-curated
02
Structure

Chai-1 / Chai-2

Chai Discovery
17 match signals

Generate or select a structural hypothesis for the target or complex.

Acceptsprotein sequencedna rna sequencesmall molecule
→
Producescomplex structure
  • Fits the small molecule modality.
  • Fits the end to end objective.
  • Matched capabilities: discovery.
  • Accepts at least one available project data type directly.
  • Has a normalized benchmark claim in BioAtlas.
Evidence 4/7 documented · benchmark curated
03
Small-molecule chemistry

DiffDock

MIT (Barzilay & Jaakkola labs)
11 match signals

Generate, dock, score or prioritize small-molecule candidates.

Acceptsprotein structuresmall moleculemolecule library
→
Producesdocking poses
  • Fits the small molecule modality.
  • Fits the end to end objective.
  • Has a normalized benchmark claim in BioAtlas.
Cautions
  • No direct declared input match for the selected available data.
Evidence 4/7 documented · benchmark curated
04
Cellular validation

scGPT

University of Toronto (Bo Wang Lab)
11 match signals

Predict or interpret cellular response and perturbation evidence.

Acceptsomics matrix
→
Producescell statemultimodal evidence
  • Fits the small molecule modality.
  • Fits the end to end objective.
  • Has a normalized benchmark claim in BioAtlas.
Cautions
  • No direct declared input match for the selected available data.
Evidence 4/7 documented · benchmark curated
05
Translation and portfolio decision

Generate:Biomedicines

Generate:Biomedicines
13 match signals

Integrate evidence into a traceable candidate or programme recommendation.

Acceptsdisease contexttarget hypothesisprotein sequencedna rna sequence
→
Producesprogram recommendationmultimodal evidencecandidate moleculesdesigned protein
  • Fits the small molecule modality.
  • Fits the end to end objective.
  • Matched capabilities: platform, discovery, pipeline.
  • Accepts at least one available project data type directly.
Cautions
  • No normalized benchmark claim has been curated yet.
Evidence 1/7 documented · benchmark not-curated
Typed handoffs

Can the selected tools exchange usable outputs?

target → structureadapter required

No direct type match; 1 explicit transformation is available.

Resolve the target hypothesis to a sequence identifier and canonical isoform.
structure → chemistryincompatible

No documented direct handoff or adapter connects these model contracts.

chemistry → cellincompatible

No documented direct handoff or adapter connects these model contracts.

cell → translationcompatible

Direct handoff available through Predicted or measured cell state, Multimodal evidence bundle.

Missing evidence

What must be resolved

  • A specific target or target class was not supplied.
  • A disease or therapeutic context was not supplied.
  • At least one selected model lacks a normalized benchmark claim in BioAtlas.
  • The generated stack contains at least one unresolved typed handoff or required-stage issue.
Validation plan

Experiments and checks

  1. Lock the exact model version, checkpoint, input preparation and inference settings.
  2. Evaluate on a project-relevant holdout set that is separated by time, scaffold, target family or biological context.
  3. Compare against a transparent baseline and record failure cases rather than reporting only aggregate performance.
  4. Confirm top candidates with orthogonal biochemical, biophysical and cellular assays; pose and affinity predictions are not substitutes for measurement.
Alternatives

Other catalogue matches

AlphaFold 2 / 3

Google DeepMind · 16 signals

Fits the small molecule modality.

Boltz-1 / Boltz-2

MIT (Barzilay & Jaakkola labs) · 16 signals

Fits the small molecule modality.

HelixFold3

PaddleHelix / Baidu · 12 signals

Fits the small molecule modality.

Protenix

ByteDance · 12 signals

Fits the small molecule modality.