A multi-omic machine-learning framework for disease classification

From the age of Hippocrates, medicine has strived to understand the determinants behind human illness and to develop effective therapeutic interventions. Over the last decade, advances in quantitative high-throughput methods have provided a detailed characterization of disease-associated perturbations and has revolutionized clinical practice. Genomics, for example, have revealed the consequences of mutations at a single-base resolution, whereas metabolomics has uncovered the role of small molecules in the blood and across the gastrointestinal tract. Similarly, microbiome research and metagenomics have exposed compositional and functional modulation of gut-dwelling microbes in cancer, type-2 diabetes (T2D), and inflammatory bowel disease (IBD). However, despite these advances, most studies to date have focused on a single disease and often based a specific molecular approach, offering a partial perspective of underlying disease mechanisms and potentially failing to capture a complete, multi-omic view. 
Recently, large scale initiatives, such as the 500 Functional Genomes (500FG) and the Integrative Human Microbiome Project (iHMP) have adopted a more principled approach, providing both deep phenotyping and multi-omic data. Together, such studies provide a more detailed and nuanced view of the underlying mechanisms of disease. Yet, due to the substantial cost of the multi-omic assays, such studies generally focus on a single or a few diseases, hindering a comparison between trajectories of similar pathologies and co-morbidities. Moreover, from a computational perspective, integrative analysis of multi-layered omic data is highly challenging, owing to the number of features involved and the complex interactions between them. 
To address these gaps, we propose to develop both model-based and machine-learning based multi-omics integration approaches that could uncover the trajectories leading from health to disease conditions. By integrating data from multiple sources, we expect to gain system-level insights into the mechanisms of disease and uncover interactions between pathologies. Our proposed methods will provide a more comprehensive and holistic view of the complex web of interactions between rich phenotypic information and human health and disease.