Integration of Multi-Omics Data Using Probabilistic Graph Models and External Knowledge

Background: High-throughput sequencing technologies have revolutionized the ability to perform systems-level biology and elucidate molecular mechanisms of disease through the comprehensive characterization of different layers of biological information. Integration of these heterogeneous layers can provide insight into the underlying biology but is challenged by modeling complex interactions. Objective: We introduce OBaNK: omics integration using Bayesian networks and external knowledge, an algorithm to model interactions between heterogeneous high-dimensional biological data to elucidate complex functional clusters and emergent relationships associated with an observed phenotype. Method: Using Bayesian network learning, we modeled the statistical dependencies and interactions between lipidomics, proteomics, and metabolomics data. The strength of a learned interaction between molecules was altered based on external knowledge. Results : Networks learned from synthetic datasets based on real pathways achieved an average area under the curve score of ~0.85, an improvement of ~0.23 from baseline methods. When applied to real multi-omics data collected during pregnancy, five distinct functional networks of heterogeneous biological data were identified, and the results were compared to other multi-omics integration approaches. Conclusion: OBaNK successfully improved the accuracy of learning interaction networks from data integrating external knowledge, identified heterogeneous functional networks from real data, and suggested potential novel interactions associated with the phenotype. These findings can guide future hypothesis generation. OBaNK source code is available at: https://github.com/bridgettripp/OBaNK.git, and a graphical user interface is available at: http://otulab.unl.edu/OBaNK.