Reinforced Active Learning for Large-Scale Virtual Screening with Learnable Policy Model

Overview

we introduce GLARE, a GRPO-based Learning framework for Active REinforced screening, designed to overcome the limitations of traditional active learning methods and enhance large-scale virtual screening. GLARE reformulates the virtual screening process as a Markov Decision Process (MDP), enabling reinforcement learning to dynamically optimize molecular selection strategies. By leveraging Group Relative Policy Optimization (GRPO), GLARE eliminates the reliance on manually-designed heuristics, learning to adaptively screen large-scale chemical spaces.

Environment

Required dependencies and versions:

• Python 3.9
• Torch 1.13.1
• Torch Geometric 2.4.0
• Numpy 1.24.0
• Pandas 2.0.3
• Scipy 1.13.1
• Scikit-learn 1.3.0
• Rdkit 2023.3.2

Data

Run utils/preprocess_data.py to preprocess ALDH1, PKM2 and VDR.

Run utils/preprocess_data_enamine.py to preprocess Enamine50k and EnamineHTS.

Training

For LIT-PCBA, run this command:

python main.py -cuda $cuda -output_folder "result_VDR" -mode "a" -architecture "ginl" -strategy "grpo" -dataset "VDR" -seed 0 -start_active_num 1 -start_num 64 -batch_size 64 -max_screen_size 1000 -ensemble_size 10 -epochs 50

For Enamine, run this command:

python main.py -cuda $cuda -output_folder "result_Enamine50k" -mode "a" -architecture "gine" -strategy "grpo" -dataset "Enamine50k" -seed 0 -start_active_num 1 -start_num 500 -batch_size 500 -max_screen_size 3000 -ensemble_size 10 -epochs 3

Citation and Contact

If you find GLARE useful for your research and applications, please cite:

@inproceedings{
    chen2025reinforced,
    title={Reinforced Active Learning for Large-Scale Virtual Screening with Learnable Policy Model},
    author={Yicong Chen and Jiahua Rao and Jiancong Xie and Dahao Xu and Zhen WANG and Yuedong Yang},
    booktitle={Thirty-ninth Annual Conference on Neural Information Processing Systems},
    year={2025},
    url={https://neurips.cc/virtual/2025/poster/119971}
}

Please contact Jiahua Rao for any questions or suggestions.