Evidence Comparison

RFdiffusion vs ProteinMPNN: Protein Design Roles

Why are RFdiffusion and ProteinMPNN usually different stages of one protein-design workflow rather than direct substitutes?

BioAtlas does not create a universal winner. The useful choice depends on scientific task, inputs, access, evaluation protocol, context of use, uncertainty and required validation.

Protein & Binder Design

RFdiffusion

Diffusion models that hallucinate brand-new proteins.

OrganizationInstitute for Protein Design, UW

AccessOpen source

DeploymentSelf-hosted

Evidence coverage5/7 evidence fields documented

BenchmarkExperimental binder validation

Metric / taskDe novo protein design

Inputs

Motif, target, symmetry or design constraints

Outputs

Novel protein backbones · Designed scaffolds / binders / enzymes

Known limitations

Performance depends on the evaluation dataset and operating conditions.
Task-specific benchmark results should not be compared across unlike domains.
Outputs require task-specific scientific and experimental validation.

Open full passport

Protein & Binder Design

ProteinMPNN / LigandMPNN

Given a shape, design the sequence that folds into it.

OrganizationInstitute for Protein Design, UW

AccessOpen source

DeploymentSelf-hosted

Evidence coverage3/7 evidence fields documented

BenchmarkNot yet curated

Metric / taskNot yet extracted

Inputs

Protein backbone · Optional non-protein atomic context

Outputs

Protein sequences

Known limitations

Performance depends on the evaluation dataset and operating conditions.
A structured benchmark claim has not yet been extracted for this record.
Outputs require task-specific scientific and experimental validation.

Open full passport

Benchmark comparability · none

Not directly comparable

One or both benchmark claims are not curated.

A direct comparison requires aligned task, dataset, metric and split/protocol context. Otherwise BioAtlas treats the evidence as partial or contextual rather than manufacturing a winner.

RFdiffusion claim →

Decision rules

What should determine the choice?

Scientific task

Confirm that both systems are being evaluated for the same task. Structure, pose, affinity and design claims are not interchangeable.

Protocol comparability

Only compare benchmark results when dataset, split, metric and implementation conditions align closely enough to support the comparison.

Access and reproducibility

Code, weights, API access, commercial terms and deployment constraints can materially change whether a model is usable in a governed programme.

Validation plan

Use the comparison to design the next validation step—not as a substitute for prospective evaluation on your own scientific problem.

Open interactive comparison Evidence methodology

Why?

Why might one model be chosen over the other?

Why this model?

Choose the model whose task, access constraints, evidence and deployment fit the actual scientific decision—not the one with the most impressive headline metric.

Why not the alternative?

A model can be scientifically strong yet inappropriate when its benchmark context, licensing, inputs, reproducibility or validation burden does not match your programme.

What evidence is missing?

If benchmark comparability is partial or contextual, the next step should be a matched evaluation on the same data, protocol and decision-relevant endpoint.