ProteinMPNN

A foundational structure-conditioned sequence-design model with strong original experimental validation across multiple protein architectures. Its residue-graph encoder and arbitrary-order decoder became reusable design components, with explicit follow-on adoption in LigandMPNN and BindCraft. This selection reflects influence and experimental grounding, not universal superiority or sole credit for downstream pipelines' results.

Chapters

1. Describe the backbone

   Backbone geometry becomes a residue-neighbor graph with atom-pair distance and chain-position features.

   Construct backbone-neighbor relations and residue-edge features.

   • Masked batched kNN graphs and neighbor gather
   • Chain-Aware Relative Position Buckets
   • Protein backbone edge geometry and 25 RBF blocks

2. Encode structural context

   Joint node/edge message updates summarize the backbone environment.

   Update residue and edge states from the local backbone graph.

   • ProteinMPNN node-and-edge message-passing layer

3. Decode in a chosen order

   Visibility masks expose earlier sequence identities; the decoder and output projection produce amino-acid logits.

   Enforce a chosen autoregressive order while decoding residue identities.

   • Arbitrary-order autoregressive graph mask
   • Sequence-conditioned ProteinMPNN decoder stack
   • Affine Projection

4. Sample or score sequences

   Constrained sampling feeds chosen identities back to the decoder. A target sequence instead supports likelihood training/scoring.

   Sample permitted identities or score a supplied target sequence.

   • Constrained categorical sampling with tied positions
   • Categorical Cross-Entropy

All exercises in this guide

• Masked Sum
• Masked Mean
• Stable vector norms, normalization, dot and cross features
• Masked pairwise distances and hard distance-bin targets
• Masked batched kNN graphs and neighbor gather
• Token Embedding Lookup
• Chain-Aware Relative Position Buckets
• Affine Projection
• Layer Normalization
• Chain, entity and token-relative position buckets
• Protein backbone edge geometry and 25 RBF blocks
• Affine projection and embedding lookup
• LayerNorm and RMSNorm from reductions
• Feedforward Block
• ProteinMPNN node-and-edge message-passing layer
• Arbitrary-order autoregressive graph mask
• Sequence-conditioned ProteinMPNN decoder stack
• Masked Softmax
• Constrained categorical sampling with tied positions
• Masked Log-Sum-Exp
• Categorical Cross-Entropy
• Stable categorical and binary classification losses