Multi-omics integration is a powerful approach to uncovering complex biological mechanisms in human disease research. This thesis evaluates and compares four multi-omics integration methods—MOFA+ (Multi-Omics Factor Analysis Plus), DIABLO (Data Integration Analysis for Biomarker discovery using Latent cOmponents), MoGCN (Multi-omics Graph Convolutional Network), and MOGONET (Multi-Omics Graph cONvolutional nETwork) using glioblastoma and breast cancer datasets. These datasets incorporate RNA expression, DNA methylation, proteomics, and somatic copy number variations to explore their complementary strengths across unsupervised, supervised, and graph-based learning paradigms.
MOFA+ revealed latent factors that explained significant variance across omics layers, identifying key pathways such as the Hippo signaling pathway and cellular metabolism. DIABLO excelled in classification and feature selection, achieving high accuracy and balanced error rates, while identifying biologically relevant features and cross-block correlations. MoGCN leveraged patient similarity networks, achieving robust classification performance and identifying biologically meaningful genes such as PAX3 and ERBB2. MOGONET used graph convolutional networks for semi-supervised learning, achieving strong accuracy and feature importance rankings, highlighting genes like GBP1 in glioblastoma and ZNF619 in breast cancer.
Each method demonstrated unique strengths: MOFA+ in unsupervised exploration, DIABLO in supervised classification, and MoGCN and MOGONET in graph-based learning. Challenges in data heterogeneity, interpretability, and computational demands remain critical barriers to clinical translation. Future directions include integrating spatial and single-cell omics, improving computational scalability, and fostering interdisciplinary collaborations to enhance the clinical impact of multi-omics research. My thesis underscores that combining diverse integration approaches provides complementary insights into cancer biology, advancing precision medicine through comprehensive multi-omics exploration.