When perfect is a warning: explanations expose shortcuts that accuracy hides in cancer gene-expression classifiers
- 1Atlantic Technological University
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.
Related
- Early vs late fusion of qMIDS with machine learning based classifiers for oral cancer diagnosis from gene expression data
- Dynamical stability and accuracy of moment closure approximations for bursty gene expression with nonlinear degradation
- Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text
- Analysis and mitigation of equipment-induced shortcuts in AI models for laparoscopic cholecystectomy
BibTeX
@misc{ayaz2026when,
title = {{When perfect is a warning: explanations expose shortcuts that accuracy hides in cancer gene-expression classifiers}},
author = {H. Ayaz and P. Verma and M. M. Ali and O. OLADIMEJI and A. O. Ibitoye},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.02.763649},
url = {https://www.biorxiv.org/content/10.64898/2026.11.02.763649v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - When perfect is a warning: explanations expose shortcuts that accuracy hides in cancer gene-expression classifiers
AU - H. Ayaz
AU - P. Verma
AU - M. M. Ali
AU - O. OLADIMEJI
AU - A. O. Ibitoye
PY - 2026
DA - 2026/11/02
PB - bioRxiv
DO - 10.64898/2026.11.02.763649
UR - https://www.biorxiv.org/content/10.64898/2026.11.02.763649v1
AB - 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.
ER -