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

Abstract

When the classes in a training set come from different data sources, a classifier can score near-perfectly while learning where a sample came from. We show this for cancer classifiers trained on tumours from The Cancer Genome Atlas and normal tissue from the Genotype Tissue Expression project. Six model families, including a tabular foundation model, reached an area under the curve of 1.000 yet called 80 to 97% of tumour adjacent normal tissue cancer, and 100% in an independent cohort; a few matched normal samples per tissue removed most of this shortcut. Near-perfect accuracy could not tell a shortcut from an honestly easy task, flagging 87 to 92% of models trained on such tasks. We introduce a label free score, computed from model explanations, that uses the fact that a source difference looks the same in every tissue whereas cancer changes do not. It raised no false alarms, ranked models by their real reliance on the source, transferred to unseen model types and data, and failed where its theory predicts.