Hypothesis
Deep learning models trained on tissue slides can infer the status of individual molecular biomarkers well enough to serve as a cheap, fast substitute for molecular testing.
What failed
Across several cancer types, datasets and modelling approaches, training data showed strong dependencies between biomarkers and clinicopathological features, so models learned confounded signals: accuracy swung with the status of co-dependent biomarkers and clinical variables, and for several biomarkers the gain over what a pathologist infers from grade was modest.
Broken assumption
A model that predicts a biomarker from histology is detecting that biomarker's own morphological signature.
Why
Correlated biomarkers and clinical features in the training sets stop models isolating any single biomarker's morphological effect; the authors argue causal rather than correlational learning is needed.
From the published paper: Dawood et al. Nature Biomedical Engineering, 2026. ↗CC BY