One molecular question. Many AI worldviews.

Best role: A single accessible route for complex prediction plus affinity-oriented outputs.

Boundary: Predicted binding metrics are model evidence, not measured affinity and not a substitute for orthogonal physics or assay data.

Best role: Integrated hit-discovery workflow tied to the Boltz structure/affinity family.

Boundary: Experimental hit campaigns are not equivalent to prospective performance on a new target; target-specific validation is required.

Best role: End-to-end protein design with dedicated interaction scoring in the same product family.

Boundary: Generation and ranking should remain separate evidence objects; model ranking is not measured binding.

Best role: Latest Chai platform generation for partner molecular-design workflows.

Boundary: Public technical detail is more limited than for open models; BioAtlas should separate platform claims from independently reproducible evidence.

Best role: Open model family spanning structure prediction and newer biomolecular design workflows.

Boundary: Developer-reported design hit rates and benchmark gains are task-specific and should remain attached to their exact protocols.

Best role: Zero-shot biologics design with controllable antibody formats and target context.

Boundary: Generative-design success claims are not interchangeable with structure-prediction benchmarks; experimental validation remains decisive.

Best role: Atom-level conditional design for binders, enzymes and nucleic-acid interfaces.

Boundary: RFD3 is a structural generator; sequence redesign and downstream filtering may still require MPNN-style methods and independent validation.

Best role: Broad complex-structure baseline and ecosystem reference point.

Boundary: Structure confidence is not binding affinity, potency, selectivity or experimental proof.

Best role: Independent learned docking evidence for pose convergence studies.

Boundary: Pose convergence across learned dockers is stronger evidence than one pose, but still not experimental confirmation.

Best role: Independent complex-prediction challenger with a distinct model lineage.

Boundary: Chai-1 prediction evidence should not be conflated with Chai-2/3 generative-design claims.

Best role: Sequence optimization when the target design is explicitly conditioned on non-protein context.

Boundary: Use alongside structural and biophysical checks rather than treating sequence likelihood as functional proof.

Best role: Independent learned complex-prediction evidence outside the AF3/Boltz/Chai lineages.

Boundary: Pose plausibility does not establish biochemical affinity or medicinal-chemistry quality.

Best role: A structurally distinct all-atom lineage useful as an independent challenger.

Boundary: Do not treat cross-model agreement as independence if models share training data, templates or related structural priors.

Best role: Geometric surface-aware alternative for challenging docking comparisons.

Boundary: Use as a complementary pose hypothesis generator, not an isolated affinity oracle.

Best role: Fast learned pose generation as an orthogonal challenger to classical docking.

Boundary: Docking confidence is not a free-energy estimate and should not be used alone to rank biological potency.

Best role: Efficient sequence design after a structural generator proposes a backbone.

Boundary: Sequence recovery and structural compatibility do not establish expression, stability, function or therapeutic developability.