Antibody library design with Transformer and Deep Reinforcement Learning

This is the source code for AB-Gen, a tool for antibody library design. It can efficiently explore the antibody space and design CDRH3 sequences that fulfill multi-property constraints.

Installation

Clone the repository

git clone [email protected]:charlesxu90/ab-gen.git
cd ab-gen

Create environment

conda env create -f environment.yml
conda activate ab-env

Download dataset and pretrained models

git lfs pull

Install netMHCIIpan

1. Download NetMHCIIpan 4.1 as shown in the guide.
2. Modify NMHOME in /netMHCIIpan-4.1/netMHCIIpan file to be your unzipped folder, e.g.

setenv	NMHOME	~/Desktop/netMHCIIpan-4.1

Test if netMHCIIpan-4.1 works properly:

python -m agent.scoring.MHCAffinity

Running the code

Generate from pretrained models

# Prior model
python generate.py --model_path data/models/Prior_10_1.704.pt --out_file result/prior/10k_samples.txt

# Agent HER2 model
python generate.py --model_path data/models/Agent_her2_final.pt --out_file result/agent/agent_oas_her2/10k_samples.txt

# Agent MPO model
python generate.py --model_path data/models/Agent_mpo_final.pt --out_file result/agent/agent_oas_mpo/10k_samples.txt

Reproduce the figures

Run jupyter-lab in command line. Then open each jupyter notebook to reproduce the figures.

License

This code is licensed under MIT License.