A deep network embedding and robust statistical framework for multi-omics integration in rare diseases: a case study in PMM2-CDG
Abstract
Integrative analysis of multiomic datasets in rare diseases is often constrained by small-n cohorts, impeding the identification of robust and biologically meaningful disease mechanisms. We developed an end-to-end framework that combines multiomic integration, robust statistical inference and network representation learning to extract clinically relevant molecular signatures from limited sample sizes. The framework prioritizes informative omic layers using Multiple Factor Analysis, identifies stable severity-associated features through stability selection and projects these features onto the human interactome using deep network embeddings to uncover functionally coherent molecular modules. We demonstrate the utility of this approach in PMM2-CDG, an ultra-rare congenital disorder of glycosylation, by integrating transcriptomic, miRNA, metabolomic and phenotypic data from ten patient-derived fibroblast cell lines spanning the clinical severity spectrum. Network modules derived from the severity-associated dimension showed significant separation between severe and mild phenotypes and revealed biological processes related to proteostasis, neurodevelopmental signaling and extracellular matrix organization. These results show that combining multiomic integration with network embedding can help address the key challenges of rare disease research, providing a computationally scalable systems biology framework for the discovery of clinically relevant disease mechanisms in small and heterogeneous cohorts.