A multi-representation deep-learning framework for accurate multicancer classification.

Area of Science:

• Oncology
• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Accurate multicancer classification is crucial for diagnosis, treatment, and prognosis in oncology.
• Existing methods often focus on limited cancer types and single genomic data modalities.
• There is a need for advanced frameworks that integrate diverse genomic features for comprehensive cancer classification.

Purpose of the Study:

• To develop and evaluate GraphVar, a novel deep learning framework for multicancer classification.
• To integrate complementary, mutation-derived imaging and numeric genomic features.
• To advance cancer classification by leveraging multi-representation learning.

Main Methods:

• GraphVar utilizes a multi-representation deep learning approach, integrating spatial variant maps (imaging) and numeric genomic features.
• A ResNet-18 backbone extracts image features, a Transformer encoder processes numeric profiles, and a fusion module combines modalities.
• Interpretability is assessed using Grad-CAM, and functional relevance is validated via KEGG pathway enrichment analysis.

Main Results:

• GraphVar achieved high performance in a cohort of 10,112 patients across 33 cancer types, with precision, recall, F1-score, and accuracy all exceeding 99.8%.
• Grad-CAM analysis demonstrated the model's ability to pinpoint biologically relevant genomic patterns.
• KEGG pathway analysis confirmed the biological significance of GraphVar-identified genes in specific cancer types (KIRC, BRCA).

Conclusions:

• GraphVar is a robust and interpretable framework for accurate multicancer classification.
• The model's high performance and ability to identify functionally relevant genomic signatures support its potential for precision diagnostics.
• Further translational studies are warranted to explore GraphVar's clinical utility in guiding therapeutic strategies.