CoDesign: Protein-Ligand Co-Design

Protein sequence-structure co-design models for enzyme engineering.

Models

DISCO

Diffusion-based Structure and Sequence Co-design model. Based on AlphaFold3 architecture with three key innovations:

• JointDiffusionModule for simultaneous sequence and structure prediction
• Frozen DPLM-650M language model for evolutionary priors
• Recycling mechanism with LigandMPNN structure encoding

EnzyGen2

ESM2 + EGNN interleaved architecture for enzyme design:

• 33-layer ESM2 Transformer with 3 EGNN layers interleaved
• NCBI taxonomy embedding for species-aware design
• SubstrateEGNN for ligand binding prediction
• Three-stage training: MLM → Motif → Full

Documentation

• docs/DISCO_v2_training_report.md - DISCO training results
• docs/DISCO_训练全流程详解.md - DISCO training guide (Chinese)
• docs/RFD3_training_report.md - RFD3 training results
• docs/RFD3_训练全流程详解.md - RFD3 training guide (Chinese)
• DISCO/DISCO_完全初学者指南.md - DISCO complete beginner's guide (9400+ lines)
• EnzyGen2/EnzyGen2_BEGINNER_COMPLETE_CN.md - EnzyGen2 complete beginner's guide (8000+ lines)

Data Processing Scripts

The scripts/ directory contains the Prot2Text enrichment pipeline:

• build_enriched_base.py - Layer 1: enzyme annotation enrichment
• compute_structure_features.py - Layer 2: structural feature extraction
• generate_route_a_constraints.py - Layer 3: constraint extraction
• build_training_subsets.py - Generate focused training subsets
• build_conditioned_csvs.py - Build metadata-conditioned CSVs
• export_prot2text_rfd3_dataset.py - Export RFD3-ready datasets