Classifying tumors from broad category to hierarchical molecular subtype: an interpretable hierarchical framework for precision oncology
- 1Tampere University
Abstract
Precise tumor subtype classification is important for guiding personalized treatment, characterizing tumor heterogeneity, and investigating molecular mechanisms. However, most existing multi-label classification models use flat architectures and do not fully exploit the inherent hierarchical relationships and biological constraints among tumor types, which limits their performance and interpretability in complex, fine-grained pan-cancer classification. Here, we present the cancer Hallmark-based Hierarchical Recursive Model (HHRM) for pan-cancer hierarchical multi-label classification. HHRM integrates hierarchical label structures with recursive neural networks, enabling simultaneous modeling of multi-level dependencies from broad cancer categories to specific subtypes. We trained the model on a comprehensive pan-cancer cohort of 11,476 samples spanning 26 primary cancer types and 405 subtypes across up to eight hierarchical levels. HHRM achieved an accuracy of 0.7924 for predicting hierarchical tumor subtypes and was further evaluated on a validation dataset of 887 samples. Compared with other hierarchical modeling frameworks, HHRM achieved higher accuracy in all three output modes. Ablation experiments showed that replacing the hallmark-based graph convolutional network reduced model performance. On an independent dataset of 1,608 breast cancer samples, HHRM also achieved an accuracy greater than 0.86. Finally, we combined SHAP with a network-based method to rank and analyze predictive genes at different levels of the tumor hierarchy, which identified both known and potentially novel biomarkers for cancer diagnosis. In conclusion, to our knowledge, HHRM is the first hierarchical model for multi-level tumor subtype classification. Its competitive performance, combined with post hoc biological interpretation, suggests potential value for precision cancer diagnosis.
Related
- Genetically programmable molecular framework for constructing precision recombinant human ECM mimics
- Molecular endotyping in systemic lupus erythematosus: an integrative framework toward precision therapy
- A standardised framework for classifying estimates of reproductive isolation
- scFLAME: a unified generative model for interpretable clustering, hierarchical structure discovery and marker-gene identification in single-cell RNA-seq data
- A Single-Cell Framework for Classifying Human Th17 Pathogenicity Links Acylcarnitine Metabolism to Non-Pathogenic Inflammation in Type 2 Diabetes
BibTeX
@misc{song2026classifying,
title = {{Classifying tumors from broad category to hierarchical molecular subtype: an interpretable hierarchical framework for precision oncology}},
author = {Q. Song and F. Liu and X. Lai},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.30.761789},
url = {https://www.biorxiv.org/content/10.64898/2026.10.30.761789v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Classifying tumors from broad category to hierarchical molecular subtype: an interpretable hierarchical framework for precision oncology
AU - Q. Song
AU - F. Liu
AU - X. Lai
PY - 2026
DA - 2026/10/30
PB - bioRxiv
DO - 10.64898/2026.10.30.761789
UR - https://www.biorxiv.org/content/10.64898/2026.10.30.761789v1
AB - Precise tumor subtype classification is important for guiding personalized treatment, characterizing tumor heterogeneity, and investigating molecular mechanisms. However, most existing multi-label classification models use flat architectures and do not fully exploit the inherent hierarchical relationships and biological constraints among tumor types, which limits their performance and interpretability in complex, fine-grained pan-cancer classification. Here, we present the cancer Hallmark-based Hierarchical Recursive Model (HHRM) for pan-cancer hierarchical multi-label classification. HHRM integrates hierarchical label structures with recursive neural networks, enabling simultaneous modeling of multi-level dependencies from broad cancer categories to specific subtypes. We trained the model on a comprehensive pan-cancer cohort of 11,476 samples spanning 26 primary cancer types and 405 subtypes across up to eight hierarchical levels. HHRM achieved an accuracy of 0.7924 for predicting hierarchical tumor subtypes and was further evaluated on a validation dataset of 887 samples. Compared with other hierarchical modeling frameworks, HHRM achieved higher accuracy in all three output modes. Ablation experiments showed that replacing the hallmark-based graph convolutional network reduced model performance. On an independent dataset of 1,608 breast cancer samples, HHRM also achieved an accuracy greater than 0.86. Finally, we combined SHAP with a network-based method to rank and analyze predictive genes at different levels of the tumor hierarchy, which identified both known and potentially novel biomarkers for cancer diagnosis. In conclusion, to our knowledge, HHRM is the first hierarchical model for multi-level tumor subtype classification. Its competitive performance, combined with post hoc biological interpretation, suggests potential value for precision cancer diagnosis.
ER -