Training distribution determines the ceiling of drug-blind cancer sensitivity prediction
A new research paper published on arXiv suggests that the current methods for predicting cancer drug sensitivity are flawed. The standard benchmark metric, global Pearson r, is misleading because it is heavily influenced by differences in drug potency rather than a model's ability to predict sensitivity for a specific tumor. When a more appropriate metric, per-drug Pearson r, is used, current drug encoding methods show no improvement over cell-only features. The study proposes that stratifying training data by mechanism-of-action can significantly improve prediction accuracy for targeted kinase inhibitors. AI
IMPACT Identifies a critical flaw in a common AI benchmark, potentially redirecting research efforts in precision oncology.