DrugInteractionML

Drug-Drug Interaction Prediction

An XGBoost pipeline that predicts adverse drug-drug interactions and their severity from molecular fingerprints, co-prescription patterns, and Snowflake-derived patient features.

Methods & contributions

• Molecular features — RDKit descriptors (MolWt, LogP, TPSA, H-bond donors/acceptors, Morgan fingerprints) with Tanimoto/Dice similarity between drug pairs.
• Severity modelling — multi-class interaction severity with SHAP explanations for each prediction.
• Signal features — Snowflake SQL derives proportional reporting ratios and co-prescription overlap windows.
• MLOps — MLflow experiment tracking, Step Functions orchestration, and drift-triggered retraining.

Data flow

drug pair → molecular FPs + patient features (Snowflake)
          → XGBoost (interaction + severity) → SHAP attribution

Technologies