InterFeat: a pipeline for finding interesting scientific features.

Finding interesting phenomena is the core of scientific discovery, but the notion of interestingness is vaguely defined and heavily reliant on manual judgment. We present InterFeat, an integrative pipeline for automating the discovery and ranking of interesting features (InterFeat) in structured biomedical data. The pipeline combines machine learning, knowledge graphs, literature search and large language models. We formalize “interestingness” as a combination of novelty, utility and plausibility. In a time-split evaluation, InterFeat was trained only on historical data, and managed to surface risk factors years ahead of their eventual discovery. Across eight major diseases, up to 21% of suggested factors appeared in the literature after the time cut-off. In a human evaluation, four senior physicians annotated InterFeat’s suggestions, deeming 28% of them interesting. Out of highly-ranked candidates, 40–53% were interesting, vs. 0–20% for SHAP and L1 baselines. InterFeat addresses the challenge of operationalizing “interestingness” scalably for any target with existing literature. Code and data: https://github.com/LinialLab/InterFeat