When perfect is a warning: explanations expose shortcuts that accuracy hides in cancer gene-expression classifiers

Cancer gene-expression classifiers trained on TCGA tumors versus GTEx normal tissue can reach AUC 1.000 while largely labeling tumor-adjacent normal as cancer (80–97%), revealing a source-data shortcut.

• Why it matters: Near-perfect accuracy can mask whether a model learns biology or simply detects which dataset a sample came from, undermining trust in cancer classification.
• What they did: Six model families, including a tabular foundation model, were trained on TCGA/GTEx; adding a few matched normal samples per tissue removed most shortcut behavior, and a label-free explanation score was derived to detect source-linked patterns.
• The result: The explanation-based score raised no false alarms, ranked models by true source reliance, transferred across unseen model types and data, and failed where its theory predicted, flagging 87–92% of such shortcut-prone tasks.

AI summary of the abstract · may contain errors · not medical advice · read the paper.