Project · 2025
Teaching models to find new drugs.
Machine learning research ranking drug and disease associations across large biomedical datasets. I engineered the feature pipelines, trained gradient-boosted rankers (XGBoost, CatBoost, LightGBM) and PyTorch nets, and validated with Top-K metrics.
What it is
Drug repurposing asks whether an existing, already-approved drug can treat a disease it was not designed for. This research project applies machine learning to that question: ranking candidate drug and disease associations across large biomedical datasets so the most promising pairs surface first.
My contribution
I engineered the feature pipelines that turn raw biomedical data into model-ready features, and I trained the ranking models: gradient-boosted rankers with XGBoost, CatBoost, and LightGBM alongside PyTorch neural networks.
Because the output is a ranked list, evaluation used Top-K metrics: what matters is whether the true associations appear near the top of the ranking, not average accuracy across every pair.
More projects
- Local RAG App: An offline retrieval-augmented generation (RAG) document chatbot built by Om Patel with Python, FastAPI, Ollama, and ChromaDB: local LLM answers with source citations and no data leaving the machine.
- Legal One: A real-time legal consultation web platform Om Patel designed, built, and deployed solo: encrypted Socket.IO messaging, role-based access for clients, lawyers, and admins, and GraphQL on Next.js, Prisma, and MongoDB.
- SW Habitation: The production marketing site and blog platform Om Patel built for the SW Habitation development studio with Next.js, TypeScript, and Tailwind CSS, with a scalable content system and performance and SEO work.