Interpretable graph Kolmogorov-Arnold networks for multi-cancer classification and biomarker identification using...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Integrating heterogeneous multi-omics data for precision cancer diagnostics presents significant analytical and computational challenges.
• Developing robust models requires effectively combining diverse data types like gene expression, microRNA, DNA methylation, and protein-protein interaction networks.
• Existing methods often struggle with data dimensionality and interpretability.

Purpose of the Study:

• To introduce Multi-Omics Graph Kolmogorov-Arnold Network (MOGKAN), a novel deep learning framework for cancer classification.
• To demonstrate MOGKAN's capability in integrating messenger RNA, microRNA, DNA methylation, and protein-protein interaction data.
• To enhance the interpretability and predictive performance of multi-omics data analysis for cancer diagnostics.

Main Methods:

• Utilized differential gene expression analysis with DESeq2, LIMMA, and LASSO regression for multi-omics data dimensionality reduction.
• Developed a deep learning framework (MOGKAN) based on the Kolmogorov-Arnold theorem principle.
• Incorporated trainable univariate functions within the MOGKAN architecture for enhanced interpretability.

Main Results:

• Achieved a high cancer classification accuracy of 96.28% across 31 cancer types.
• Demonstrated low experimental variability compared to other deep learning models.
• Identified and validated cancer-related biomarkers through Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis.

Conclusions:

• MOGKAN effectively integrates multi-omics data with graph-based deep learning for robust cancer classification.
• The framework offers significant interpretability, facilitating the translation of complex biological data into clinical applications.
• MOGKAN shows potential for advancing precision cancer diagnostics through enhanced predictive performance.