iMVAN: integrative multimodal variational autoencoder and network fusion for biomarker identification and cancer subtype classification

Abstract

Numerous research has been conducted to define the molecular and clinical aspects of various tumors from a multi-omics point of view. However, there are significant obstacles in integrating multi-omics via Machine Learning (ML) for biomarker identification and cancer subtype classification. In this research, iMVAN, an integrated Multimodal Variational Autoencoder and Network fusion, is presented for biomarker discovery and classification of cancer subtypes. First, MVAE is used on multi-omics data consisting of Copy Number Variation (CNV), mRNA, and Reverse Protein Phase Array (rppa) to discover the biomarkers associated with distinct cancer subtypes. Then, multi-omics integration is accomplished by fusing similarity networks. Ultimately, the MVAE latent data and network fusion are given to a Simplified Graph Convolutional Network (SGC) for categorizing cancer subtypes. The suggested study extracts the top 100 features, which are then submitted to the KEGG analysis and survival analysis test. The survival study identifies nine biomarkers, including AGT, CDH1, CALML5, ERBB2, CCND1, FZD6, BRAF, AR, and MSH6, as poor prognostic markers. In addition, the cancer subtypes are classified, and the performance is assessed. The experimental findings demonstrate that the iMVAN performed well, with an accuracy of 87%.