AI powered Drug Interaction Detection System (Deerhack 2026 AI/ML Track Winner)

Built a graph neural network (GraphSAGE encoder + bilinear decoder + MLP harm head) over a 21,257-node biological graph (2,135 drugs + 19,122 proteins, 1.8M edges) to predict drug-drug interactions never explicitly tested Trained on 4.6M DECAGON polypharmacy records (63,473 drug pairs, 1,317 side effects), achieving 0.884 ROC-AUC and 98.8% recall - tuned for high sensitivity so real interactions are never missed in a clinical safety context Deployed a FastAPI backend and React/Vite dashboard producing a per-pair harm score and combined polypharmacy risk grade (CRITICAL/HIGH/MODERATE/LOW) from any drug list Integrated an LLM pharmacology chatbot pre-loaded with detected interaction context, plus per-patient history and PDF report export