How hxyfighter/molrl-mgpt works
Project type and purpose
• This is an open-source research framework (“MolRL-MGPT”) for de novo drug design using reinforcement learning (RL) with multiple GPT-style agents.
• It provides two main workflows: (a) pre-training a molecular language model on big SMILES datasets, and (b) running multi-agent RL for goal-directed molecule generation (e.g., GuacaMol tasks or SARS-CoV-2 docking).
File structure
• Top level:
• README.md (project overview and instructions)
• codes/ (all Python modules and entry scripts)
• data/ (raw datasets, vocabulary, protein target files)
• codes/ directory:
• pretrain.py + model.py + vocabulary.py + dataset.py – the pre-training pipeline
• MARL.py + MARL_guacamol.py + assess_goal_directed_generation.py – multi-agent RL orchestration and evaluation
• scoring_function.py + utils.py – environment helpers and external docking/score wrappers
• data/ directory:
• ChEMBL_Smiles.csv, vocab.txt – SMILES corpus and token vocabulary
• targets/*.pdbqt – protein structures for docking oracles
• Key config and dependency files are implicit (no Dockerfile, but uses PyTorch, RDKit, OpenBabel, QuickVina).
README for additional insights
• Workflows:
a. Pre-training: python codes/pretrain.py on a filtered ZINC or ChEMBL SMILES dataset.
b. Multi-agent RL: python codes/MARL.py --task_id … for GuacaMol tasks or docking oracles.
• Dependency on external tools: RDKit for chemistry operations, OpenBabel/QuickVina for docking, PyTDC for property evaluation, TensorBoard for logging.