Classifying tumors from broad category to hierarchical molecular subtype: an interpretable hierarchical framework for precision oncology

The cancer Hallmark-based Hierarchical Recursive Model (HHRM) classifies pan-cancer hierarchical tumor subtypes with accuracy 0.7924 across 11,476 samples spanning 26 primary types and 405 subtypes.

• Why it matters: Flat multi-label models underuse hierarchical tumor relationships and biological constraints, limiting performance and interpretability for complex, fine-grained pan-cancer subtype classification.
• What they did: HHRM couples hierarchical label structures with recursive neural networks and trains on 11,476 samples across up to eight hierarchy levels; it is validated on 887 samples and tested on 1,608 independent breast cancer samples.
• The result: HHRM predicts hierarchical subtypes with 0.7924 accuracy, exceeds other hierarchical frameworks in all three output modes, and achieves >0.86 accuracy in breast cancer; SHAP plus a network method ranks predictive genes across hierarchy levels.

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