Drug Repurposing

Problem

Prioritize candidate drug and disease associations for further research. The task combines biomedical sources with different signals and produces a ranked shortlist for experimental follow-up.

My contribution

I unified DrugBank, CTD, and BioSNAP into more than 1.2 million drug and disease pairs, roughly 1.5 GB of data, and built the 27-feature pipeline for chemical, structural, and biological signals.

I trained and tuned XGBoost and CatBoost, and supported teammates working on LightGBM and neural-network experiments.

Technical decisions

The output is a ranked list, so evaluation used Top-K metrics with a 60/20/20 training, validation, and test split. This measures whether known associations appear near the top of the candidate list, which fits the research task better than a single classification score.

Results

XGBoost performed best in the project evaluation, reaching about 95% Top-5 accuracy and about 22% Top-1. These are ranking results on the project's biomedical dataset, not clinical validation or evidence that a candidate treatment is effective.