Train GradientBoosting on 300 drugs x 839 GEO disease signatures with
Repurposing-Hub indications as labels (432 positives), disease-grouped CV.
Finding: 0.925 CV AUC looks like a win but is a MIRAGE. Feature
importances are all drug-level (drug_std 0.33, drug_mean 0.30,
broadness 0.17); drug-disease connectivity importance = 0.01. The model
learned a drug-POPULARITY prior, not disease-specific matching. On
held-out sickle it ranks hydroxyurea 231/300 (worse than baseline) and
tops out with promiscuous drugs (dexamethasone, methotrexate). Classic
degree-bias trap. Connectivity also has ~chance AUC (0.51) for predicting
approved indications.
Both obvious approaches now fail instructively: unsupervised = specificity
ceiling; naive supervised = degree bias. Real progress needs degree-
debiased training + much larger clean labels (a research effort).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Junior B.
649f617019