OncoPath

repo

OncoPath is an AI-driven clinical decision-support platform that predicts organ-specific metastatic risk by fusing longitudinal clinical data, 101-gene mutation profiles, and tumor pathology imaging. Trained on the MSK-MET cohort across all 21 metastatic destinations, the platform lets clinicians and researchers run real-time "What-If" simulations to see how specific mutations (e.g., TP53, KRAS, HER2) shift cancer progression across the body.

Problem

Cancer prognosis hinges on a complex interplay between a patient's clinical profile, tumor genomics, and histology — but clinicians typically reason about these signals in isolation. Existing tools either ignore the genomic layer, lack visual feedback, or fail to surface the organotropism patterns (KRAS→colorectal liver spread, HER2→breast-to-liver) that drive treatment planning.

Solution

• Multimodal fusion engine — XGBoost tabular models augmented with embeddings from the Phikon pathology foundation model, plus an ensemble layer that triggers image-based reasoning when image confidence is decisive (>70% / <30%).
• 3D Metastatic HUD — Interactive Three.js viewer that renders risk intensity as a dynamic anatomical heatmap across the human body.
• OncoBot clinical assistant — Domain-restricted RAG chatbot built on Anthropic Claude that surfaces interpreted clinical insights directly in the dashboard.
• Real-time mutation toggling — Sub-50 ms inference against 21 organ-specific models via a FastAPI backend, so users instantly see the impact of flipping a mutation on or off.
• Temporal simulation — Gompertz-based timeline modeling of metastatic spread with treatment-adjusted modes (chemotherapy, oral, etc.).
• Clinical audit trail — SHAP interpretability plus a "Genomic Lift" report that validates predictions against published organotropism patterns.

Tech Stack

Frontend — Next.js 15 (App Router) + TypeScript + React 19; Three.js, React Three Fiber, React Three Drei for the 3D viewer; Tailwind CSS v4, Radix UI, lucide-react, Framer Motion; Zustand for client state; Clerk for auth; Anthropic SDK for streaming chat.

Backend & AI — FastAPI + Uvicorn inference service; XGBoost + Scikit-learn for 21 organ-specific classifiers; PyTorch + HuggingFace Transformers + Phikon (pathology foundation model) for image embeddings; SHAP for model interpretability; Anthropic Claude (Haiku & Sonnet) for the RAG clinical assistant; Joblib for model serialization.

Data & Infrastructure — Supabase (PostgreSQL, auth, storage); Pandas / NumPy / SciPy for data processing; MSK-MET multi-omic dataset as the training corpus; Python 3.9+ / Node.js 18+ runtime.

Results

• 21 organ-specific XGBoost models deployed, with top performers reaching AUC 0.93 (Female Genital), 0.91 (Male Genital), 0.83 (Bladder) on the MSK-MET cohort.
• Positive Genomic Lift on the clinically complex sites where the model needs to read genuine genomic signal rather than shortcuts: Liver (+0.026), Lung (+0.025), CNS/Brain (+0.022), Bone (+0.015).
• Audited against oncological literature for known organotropism (e.g., HER2 → 8.5% Liver Risk Lift; KRAS → Colorectal patterns).
• Multimodal ensemble surfaces per-simulation visual_lift to quantify pathology-image influence on each risk estimate.