Title: Interpretable Graph Kolmogorov-Arnold Networks for Multi-Cancer Classification and Biomarker Identification using Multi-Omics Data

Abstract: A significant hurdle in the realm of precision cancer diagnostics lies in the development of analytical and computational models capable of handling heterogeneous multi-omics datasets at a systems level. To address this, we present the Multi-Omics Graph Kolmogorov-Arnold Network (MOGKAN), a novel deep learning framework. This model leverages data from messenger-RNA, micro-RNA sequences, and DNA methylation samples, alongside Protein-Protein Interaction (PPI) networks, to classify 31 distinct cancer types.

To reduce the dimensionality of multi-omics data while retaining essential biological characteristics, the proposed method integrates differential gene expression analysis with Linear Models for Microarray (LIMMA), DESeq2, and Least Absolute Shrinkage and Selection Operator (LASSO) regression. Grounded in the Kolmogorov-Arnold theorem, the model’s architecture employs trainable univariate functions to facilitate feature analysis and improve interpretability.

In performance evaluations, MOGKAN demonstrated a classification accuracy of 96.28 percent, showing lower experimental variability than comparable deep learning models. Furthermore, biomarkers identified by MOGKAN were confirmed as cancer-related through Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. By merging graph-based deep learning with multi-omics integration, this approach offers robust predictive power and transparency, highlighting its potential to transform complex multi-omics data into actionable clinical diagnostics for cancer.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC