In our previous work, we developed a tool called iOmicsPASS which searches for subnetworks of molecular interactions within and between -omics data types that can predict different phenotypic groups. The tool incorporates biological networks for the measured molecules in pursuit of the predictive interactions of each phenotypic group, and we showed that this approach is highly efficient in the integration of proteome and transcriptome data for cancer subtyping. Nonetheless, the prior information of biological network may not always be available between certain types of molecules, for instance between lipids and proteins. Hence, a more data-driven approach to predict unknown interactions would be desirable.
We predict the molecular interactions between lipids and proteins by considering the cross-covariance matrix between the two types of –omics data. By applying a shrinkage method, we de-noise the covariance matrix to obtain a sparse matrix which highlights the most strongly correlated molecule pairs, rendering a set of predicted interactions between and within the lipidome and proteome. Then, we applied iOmicsPASS, using the derived lipid-protein interactions in the blood plasma, to identify a set of predictive molecular interactions which can best differentiate individuals of high and low risk of coronary artery disease (CAD). We also explored the biological pathways which were up- and down-regulated in the individuals at a higher risk of CAD.
We illustrate iOmicsPASS using proteomics and lipidomics data from a Singapore Chinese cohort. Individuals were classified into high and low-risk groups based on Framingham risk scores and other cardiovascular risk factors, and lipid-protein and protein-protein interactions predictive of CAD risk groups were identified by iOmicsPASS analysis. We also re-categorize subjects based on their measured plaque volume index in their coronary arteries to find lipid and protein signatures which could potentially highlight early coronary plaque burden in asymptomatic individuals.