HallmarkGraph: a cancer hallmark informed graph neural network for classifying hierarchical tumor subtypes

Area of Science:

• Computational biology
• Machine learning in oncology
• Genomics and transcriptomics

Background:

• Accurate tumor subtype diagnosis is critical for precision oncology but faces challenges in model interpretability, cost, and validated pan-cancer classification.
• Existing methods struggle to balance accuracy with interpretability and lack validated models for hierarchical tumor subtype classification across diverse cancer types.

Purpose of the Study:

• To introduce HallmarkGraph, the first biologically informed graph neural network for classifying hierarchical tumor subtypes in human cancer.
• To develop a model that integrates transcriptome profiles and gene regulatory interactions for multi-label classification.
• To address the need for validated, accurate, and interpretable models in pan-cancer diagnostics.

Main Methods:

• Developed HallmarkGraph, a graph neural network integrating transcriptome data and gene regulatory interactions.
• Evaluated the model on a pan-cancer cohort of 11,476 samples across 26 cancers with 405 subtypes.
• Utilized 5-fold cross-validation and a separate validation dataset (887 samples) for performance assessment.
• Employed SHAP (SHapley Additive exPlanations) for feature attribution and linking cancer hallmarks to genes.

Main Results:

• HallmarkGraph achieved high 5-fold cross-validation accuracy (85%-99%) for tumor subtype classification.
• Demonstrated good generalizability on an independent validation dataset.
• Benchmarking revealed the significant role of the integrated multilayer perceptron in classifier accuracy.
• SHAP analysis identified key genes influencing model decisions, linking them to cancer hallmarks.

Conclusions:

• HallmarkGraph provides a biologically informed machine learning framework for tracking tumor transcriptomic trajectories.
• The model effectively distinguishes inter- and intra-tumor heterogeneity in pan-cancer analysis.
• This approach shows promise for improving the accuracy and interpretability of cancer diagnostics in precision oncology.