TRC-TopoGen is a modular Python platform for genomic-context-aware topoisomerase trapping molecule discovery. It combines compound curation, genomic feature engineering, proxy-supervised predictive modeling, conditional molecular generation, reinforcement learning, docking, ranking, and dossier export in one reproducible codebase.
- Hydra and OmegaConf driven configuration with experiment composition
- RDKit and PyTorch Geometric based chemistry and graph modeling stack
- Genomic window featurization and TRC hotspot scoring
- Proxy-supervised bias prediction for hotspot versus neutral topoisomerase trapping
- Conditional graph VAE generation and RL optimization
- Chemistry filtering, docking, ranking, and dossier generation
- Deterministic runs with metadata snapshots and checkpointing
conda env create -f environment.yml
conda activate trc_topogen
pip install -e .python scripts/build_compound_dataset.py
python scripts/preprocess_genomics.py
python scripts/build_trc_scores.py
python scripts/train_activity_predictor.py
python scripts/train_bias_predictor.py
python scripts/pretrain_generator.py
python scripts/finetune_generator.py
python scripts/run_rl_optimization.py
python scripts/generate_candidates.py
python scripts/filter_candidates.py
python scripts/rank_candidates.py
python scripts/export_dossier.pyThe repository follows the structure required by the TRC-TopoGen engineering specification, including dedicated configs/, scripts/, src/trc_topogen/, and tests/ directories.
Every run saves:
- resolved config snapshot
- metrics
- checkpoints
- package version manifest
- seed and hardware metadata
Aaryan Senthilvanan
Bias predictions are proxy supervised and intended for prioritization, not causal or clinical claims. Experimental validation is required for any generated candidate.