Computational Methods and Resources for Multi-omics Understanding of Complex Human Systems

Multi-omics data, increasingly available in epidemiological cohorts with rich phenotype data, pose grand opportunities and challenges. Harnessing the power of deep learning and multi-modal large language models holds great promises for multi-omics integration. I will first introduce our novel method OMEGA, Omics Multi-modality Embedding via Graphical & Articulated data, for multi-omics integration by simultaneously considering image, text and tabular representations of omics data. We applied our models to 17 incident clinical endpoints in 23,776 UK Biobank individuals with 159 metabolites (NMR based Nightingale platform) and 2,923 proteins (Affinity-based Olink 3k proteomics). Our OMEGA models outperformed alternatives including mixOmics and MOGONET. Spatial transcriptomics (ST) technologies have revolutionized our ability to study gene expression within the spatial context of tissues. Spatial variable genes (SVGs), revealing critical information and about tissue architecture, cellular interactions, and disease-relevant microenvironments, are of keen interest in ST studies. We performed a comprehensive benchmarking study of 20 state-of-the-art SVG detection methods using >600 human slides from STimage-1K4M, a large-scale ST resource comprising >1000 slides from >18 tissue types. We also constructed the first cross-tissue atlas of SVGs, enabling comparative analysis of spatial gene programs across cancer and normal tissues.