Model Book/ProteinMPNN

ProteinMPNN

A

UW Baker Lab

Sequence DesignInverse Foldingopen
Updated 2 months ago
NextIn take

ProteinMPNN is the quiet workhorse of the protein design revolution — 2M parameters, runs on a laptop, and has been behind more successful designed proteins than any model 100x its size.

Specifications

ArchitectureMessage-passing neural network on protein structure graphs
Parameters~2M
Training DataPDB structures with sequence-structure pairs
LicenseMIT license. Fully open weights and code.
HardwareRuns on CPU. GPU optional for speed.
Inference CostNegligible — runs on laptop hardware
API AvailableNo
Weights AvailableYes

Benchmark Performance

BenchmarkScore
ProteinGym52% recovery rate

When to Use This

  • +Designing sequences for computationally generated backbones
  • +Redesigning natural proteins for improved stability or function

When NOT to Use This

  • Generating new backbone structures — use RFdiffusion first
  • Predicting structure from sequence — this does the reverse

Production Readiness

production

Known Users

  • Baker Lab and hundreds of academic groups
  • Multiple protein engineering companies

Grade Rationale

A

The gold standard for inverse folding and sequence design. Tiny model, runs anywhere, works reliably, and has been experimentally validated hundreds of times. A because it does one thing and does it better than anything else.

Sources

Update History

2026-02-01Initial entry