Model Book/AlphaFold-Multimer

AlphaFold-Multimer

A-

Google DeepMind

Complex PredictionProtein-Protein Interactionrestricted
Updated 1 month ago
NextIn take

AlphaFold-Multimer is the reliable workhorse that most structural biology labs still use daily — it's not the frontier anymore, but the tooling ecosystem around it is unmatched.

Specifications

ArchitectureAlphaFold2 with multi-chain modifications for complex prediction
Parameters~400M
Training DataPDB multimer structures, UniRef, BFD
LicenseCode open-source (Apache 2.0) but model use restricted by DeepMind terms. Superseded by AF3 for new work.
Hardware1x A100 40GB+ for most complexes, more for large assemblies
Inference CostSelf-hosted — moderate GPU cost, slower than single-chain AF2
API AvailableYes
Weights AvailableNo

Benchmark Performance

BenchmarkScore
DockQ0.65 DockQ
CASP150.78 GDT-TS (multimer)

When to Use This

  • +Protein-protein interaction prediction with established tooling
  • +When you need community support and well-tested pipelines

When NOT to Use This

  • New projects where AF3 or Boltz-1 would give better results
  • Ligand docking — not designed for small molecules

Production Readiness

production

Known Users

  • Thousands of academic and industry labs worldwide
  • Isomorphic Labs

Grade Rationale

A-

Was the best complex prediction tool before AF3 arrived. Still reliable and battle-tested, with a larger user base and more community tooling than AF3. A- because it's proven but no longer state-of-the-art.

Sources

Update History

2026-02-10Initial entry