Decoupled Sequence and Structure Generation for Realistic Antibody Design

This is a public repository for our paper Decoupled Sequence and Structure Generation for Realistic Antibody Design.

Overview

• Environment
  • Conda
• Data Preparation
• Benchmark Experiments
  • Task 1: Sequence and Structure Modeling
  • Task 2: Antibody-binding CDR-H3 Design
• Acknowledgements
• Citation

Environmental Setup

Conda

# clone project
git clone https://github.com/nayoung10/ASSD.git
cd ASSD

# setup environment
source scripts/install.sh

Data Preparation

Download

1. Download all_structures.zip from the SAbDab download page.
2. Move all_structures.zip to ASSD/MEAN
3. Unzip with unzip all_structures.zip

Pre-processing for MEAN

We use the data-preprocessing scripts provided by MEAN.

cd MEAN # in ASSD/MEAN directory
bash scripts/prepare_data_kfold.sh summaries/sabdab_summary.tsv all_structures/imgt
bash scripts/prepare_data_rabd.sh summaries/rabd_summary.jsonl all_structures/imgt summaries/sabdab_all.json
bash scripts/prepare_data_skempi.sh summaries/skempi_v2_summary.jsonl all_structures/imgt summaries/sabdab_all.json

Pre-processing for ESM2

# copy data splits to ASSD/data
mkdir -p ../data && \
rsync -avm --include='*/' \
--include='train.json' \
--include='valid.json' \
--include='test.json' \
--include='*.pdb' \
--exclude='*' \
summaries/ ../data/

# data pre-processing for ESM2
cd ../ # in ASSD directory
bash scripts/prepare_data_kfold.sh
bash scripts/prepare_data_rabd.sh
bash scripts/prepare_data_skempi.sh

Benchmark Experiments

Task 1: Sequence and Structure Modeling

We first fine-tune/evaluate the sequence design model ESM2, then use its sequence predictions \hat{s} to train/evaluate the structure prediction model MEAN. For more information on running MEAN, please visit the MEAN github page.

# Step 1: Sequence design
bash scripts/k_fold_train.sh # training sequence design model 
bash scripts/k_fold_eval.sh # evaluate AAR
bash scripts/average_results_kfold.sh # average results across all folds

# Step 2: Structure prediction 
bash scripts/generate_seqs_kfold.sh # generate sequence -- i.e., the input of the structure prediction model 
cd MEAN # in ASSD/MEAN directory 
GPU=0 bash scripts/k_fold_train.sh summaries 111 mean 9901
GPU=0 bash scripts/k_fold_eval.sh summaries 111 mean 0

Task 2: Antibody-binding CDR-H3 Design

# Step 1: Sequence design
bash scripts/task2_train.sh # training sequence design model 
bash scripts/task2_eval.sh # evaluate AAR and CoSim

# Step 2: Structure prediction (MEAN)
bash scripts/generate_seqs_rabd.sh # generate sequence -- i.e., the input of the structure prediction model
cd MEAN # in ASSD/MEAN directory 
GPU=0 MODE=111 DATA_DIR=summaries/cdrh3 bash train.sh mean 3
GPU=0 MODE=111 DATA_DIR=summaries/cdrh3 bash rabd_test.sh 0

Acknowledgements

We deeply appreciate the following works/repositories, on which our project heavily relies.

• Structure-informed Language Models Are Protein Designers for training/testing pipelines of the sequence design model.
• MEAN: Conditional Antibody Design as 3D Equivariant Graph Translation for data-preprocessing and the entire structure prediction model.
• Evolutionary Scale Modeling as the pLM adopted for the sequence design model.
• Scoring function for automated assessment of protein structure template quality for tm-scoring of the generated structures.

Citation