Training distribution determines the ceiling of drug-blind cancer sensitivity prediction
A new study reveals that current methods for predicting cancer drug sensitivity are limited by a metric artifact rather than a lack of sophisticated drug representations. The standard benchmark, global Pearson r, is heavily influenced by general drug potency differences, which a simple drug-mean predictor can already capture. By shifting to a per-drug Pearson r metric and stratifying training data by mechanism-of-action, researchers significantly improved prediction accuracy, particularly for kinase inhibitors. AI
IMPACT Highlights how AI model training and evaluation methodologies can obscure true performance, impacting the development of AI-driven precision medicine.