94 / 2026-04-07 00:00:29

Identification and Clinical Validation of Core Invasion Risk Modules in Ovarian Cancer Based on Multimodal Deep Learning

Ovarian cancer,Multimodal deep learning,Invasion risk module,Prognostic prediction

Ovarian cancer remains the leading cause of gynecological cancer mortality due to late diagnosis and aggressive invasion. To address this, we developed a masked-aware multimodal Variational Autoencoder (mmVAE-Cox) framework to systematically identify core molecular modules driving invasion. By integrating multi-omics data (transcriptome, methylome, CNV, WSI) from TCGA-OV (n=559) and external cohorts, our model achieved superior prognostic prediction (C-index=0.714). Through feature attribution and WGCNA analysis, we pinpointed three key modules (ME1, ME4, ME12) involving genes like CDC42BPB, COL11A1, and C7, which are implicated in GTPase signaling, ECM organization, and complement activation. Furthermore, preliminary Hi-C data suggested that 3D chromatin remodeling may regulate these core genes. This study provides a robust multi-modal analytical pipeline and identifies novel molecular targets for assessing invasion risk and improving clinical management.