EVO ColabDesign

This code implements a Quality-Diversity Algorithm for binder design with ColabDesign.

Implementation Details

Archive Structure

The implementation uses a GridArchive with discrete binning optimized for protein sequences:

• Sequence length: One bin per integer length (categorical variable)
• Amino acid composition: 3D binning over hydrophobic, polar, and charged percentages (other percentage derived as 1 - sum)
• Niche count: n_lengths × (category_bins)³ — typical configurations yield 150-700 niches

Quality-Diversity Optimization

• Fitness: Negative AlphaFold loss (lower loss = higher fitness)
• Diversity: Maintained through amino acid category composition and sequence length
• Operators: Mutation (position-wise AA substitution) and crossover (chunk-based recombination with pLDDT-guided selection)

Key Implementation Notes

• All loss terms are length-normalized to prevent bias toward short/long sequences
• Model state (weights, inputs) is preserved across sequence evaluations
• Per-niche elite structures are automatically saved and updated when improved
• Full per-term loss breakdown logged in JSONL format for post-hoc analysis

Usage

Running MAP-Elites

af_model.design_mapelites(
    iters=100,
    num_elites=700,  # higher count enables finer category resolution
    mutation_rate=5,
    num_sequences=200,
    min_len=20,
    max_len=40,
    experiment_name="my_experiment",
    init_sampling_strategy="stratified"  # or "random"
)

Output:

• experiment_name/all_seq.jsonl: Full loss breakdown per evaluated sequence
• experiment_name/elite_structure/: Best PDB structure per niche (auto-updated)
• experiment_name/final_elites.jsonl: Final archive with complete metrics

OpenBLAS Configuration

If you encounter OpenBLAS errors:

set -x OPENBLAS_NUM_THREADS 1
set -x OMP_NUM_THREADS 1
set -x MKL_NUM_THREADS 1

export OPENBLAS_NUM_THREADS=1
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1

---

Original ColabDesign README

ColabDesign

Making Protein Design accessible to all via Google Colab!

• P(structure | sequence)
  • TrDesign - using TrRosetta for design
  • AfDesign - using AlphaFold for design
  • [WIP] RfDesign - using RoseTTAFold for design
• P(sequence | structure)
  • ProteinMPNN
  • [WIP] TrMRF
• P(sequence)
  • [WIP] MSA_transformer
  • [WIP] SEQ - (GREMLIN, mfDCA, arDCA, plmDCA, bmDCA, etc)
• P(structure)
  • Rfdiffusion

Where can I chat with other ColabDesign users?

• See our Discord channel!

Presentations

Slides Talk

Contributors:

• Sergey Ovchinnikov @sokrypton
• Shihao Feng @JeffSHF
• Justas Dauparas @dauparas
• Weikun.Wu @guyujun (from Levinthal.bio)
• Christopher Frank @chris-kafka