GitHub - chq1155/RL-PLM: Official implementation of From Supervision to Exploration: What Does Protein Language Model Learn During Reinforcement Learning? (original) (raw)

This repository contains the implementation of the RL-PLM framework introduced in our paper:
From Supervision to Exploration: What Does Protein Language Model Learn During Reinforcement Learning? (Arxiv preprint).

img

RL-PLM provides a unified platform for analyzing how reinforcement learning enhances and reshapes Protein Language Models (PLMs) across four biological design systems.
The framework supports multiple RL algorithms—DPO, PPO, and GRPO—and integrates them with both auto-regressive and masked-language protein models.

img
Figure 1. Conceptual analogy of RL for protein design—task difficulty, policy capacity, and reward accuracy jointly determine learning efficacy.

📂 Repository Structure


RL-PLM/
│
├── amp_design/          # Antimicrobial peptide (AMP) design via RL
├── kinase_mutation/     # Kinase optimization via sequence mutation
├── antibody_mutation/   # Antibody–antigen binding optimization
└── inverse_folding/     # Structure-conditioned sequence generation

img
Figure 2. Four biological systems implemented in RL-PLM.

📦 Data and Model Weights

All datasets and pretrained checkpoints can be downloaded here:
RL_PLM_data (Google Drive)

Unzip and place the contents under your BASE_PATH before running any experiment.

🧬 Citation

If you find this repository useful, please cite:

@article{cao2025supervision, title={From Supervision to Exploration: What Does Protein Language Model Learn During Reinforcement Learning?}, author={Cao, Hanqun and Zhang, Hongrui and Xu, Junde and Zhang, Zhou and Shen, Lingdong and Sun, Minghao and Liu, Ge and Xu, Jinbo and Li, Wu-Jun and Ni, Jinren and others}, journal={arXiv preprint arXiv:2510.01571}, year={2025} } }