MOGSP: A MULTI-OMICS GATED SPARSE PERTURBATION FRAMEWORK FOR PAN-CANCER CLASSIFICATION AND BIOMARKER DISCOVERY

Abstract

Pan-cancer classification from multi-omics data remains a central problem in computational oncology because models must integrate heterogeneous molecular layers while preserving interpretability at both modality and gene levels. We present MoGSP (Multi-Omics Gated Sparse Perturbation), an interpretable deep learning framework that jointly integrates RNA-seq expression, DNA methylation and somatic mutation profiles through modality-specific encoders, multi-head cross-modal attention, an adaptive softmax gating mechanism and a variational latent representation. Unlike static concatenation-based fusion, MoGSP learns sample-specific modality weights and couples these weights with sparse perturbation analysis to derive gene-level impact scores. Applied to 10,702 TCGA samples across 32 cancer types, MoGSP achieved 96.39% held-out test accuracy and a macro-F1 score of 0.963, outperforming RNA-only and naïve multi-omics concatenation baselines. The adaptive gate recovered a biologically coherent dominance of DNA methylation while identifying RNA-elevated and mutation-elevated patient subgroups. Sparse perturbation highlighted known cancer-associated genes, including ST14 and HOXC6, and nominated less-characterised candidates requiring independent validation. These results suggest that adaptive, interpretable multi-omics fusion can support robust pan-cancer classification and hypothesis-generating biomarker discovery.