Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design

Introduction

This repository provides the source codes associated with the paper Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design.

Environment and External Tool Setup

• Create the primary environment: conda env create -f environment.yml
• Create the environment for chemprop: conda env create -f chemprop.yml
• Install QuickVina2-GPU-2.1: https://github.com/DeltaGroupNJUPT/Vina-GPU-2.1

Data Source

Molecules (download molecular data in sdf format):

• QM9: http://deepchem.io.s3-website-us-west-1.amazonaws.com/datasets/gdb9.tar.gz
• ZINC15: https://zinc15.docking.org/tranches/home/#
• PubChem: https://pubchem.ncbi.nlm.nih.gov/#query=small%20molecule&tab=compound

Protein:

• 6VHN: https://www.rcsb.org/structure/6VHN

Uncertainty Prediction

• Install chemprop: https://github.com/chemprop/chemprop
• cd uncertainty
• Train surrogate models: ./train_surrogates.sh [datasetName] [property]

Diffusion Models

• cd scripts
• Train diffusion models: ./a_run_zz_train.sh [dataset] [removeH] [batchsize] [n_layers] [nf] [n_epoch] [CUDA_DEVICES] [properties]
• Optimize diffusion models: ./a_run_zz_optimize.sh
• Generate molecules and evaluate: ./a_run_zz_evaluate.sh [exp_name]