How edouard-rabasse/hackaton-margo works
Project type and purpose
• This is a 48-hour hackathon codebase for molecular toxicity prediction, organized by Margo, IBM, and Qubit Pharmaceuticals.
• It’s an end-to-end ML research & prototyping project that (a) ingests SMILES strings and molecular descriptors, (b) featurizes them, (c) trains multiple models (GNNs, DNNs on fingerprints/descriptors, XGBoost+Optuna), and (d) combines outputs for submission.
File structure
Top-level:
• main.py & models.py → likely entry-point orchestrators for experiments
• environment.yml & requirements.txt
• visu.py → visualization of results
• ChembertA/chemberta.py → BERT-based chemical embedding prototype
• DeepHIT/… → core GNN + DeepHit model pipelines
• edouard/xgboost_optuna.py → XGBoost with Optuna hyperparameter tuning
• simon/… → Jupyter notebooks for exploratory ML pipelines
DeepHIT folder structure shows:
• train_*.py → separate scripts for GCN, descriptor DNN, fingerprint DNN, fully-connected
• gnn.py, gnn2.py, GNN3.py → successive GNN implementations
• combine*.py → ensemble/stacking pipelines
• utils.py → shared functions
• main.py → pipeline entry
• models.py → model definitions
README insights
• Project goal: predict toxicity from SMILES + molecular features.
• Approach: represent molecular geometry via Graph Neural Networks, adapt DeepHit model.