Utilizing Pre-trained Network Medicine Models for Generating Biomarkers, Targets, Re-purposing Drugs, and Personalized Therapeutic Regimes: COVID-19 Applications

Abstract

In this paper, we present a novel pre-trained network medicine model called Selective Remodeling of Protein Networks by Chemicals (SEMO). We divide the global human protein-protein interaction (PPI) network into smaller sub-networks, and quantify the potential effects of chemicals by statistically comparing their target and non-target gene sets. By combining 9607 PPI gene sets with 2658 chemicals, we created a pre-trained pool of SEMOs, which we then used to identify SEMOs related to Covid-19 severity using DNA methylation profiling data from two clinical cohorts. The nutraceutical-derived SEMO features provided an effective model for predicting Covid-19 severity, with an AUC score of 81% in the training data and 80% in the independent validation data. Our findings suggest that Vitamin D3, Lipoic Acid, Citrulline, and Niacin, along with their associated protein networks, particularly STAT1, MMP2, CD8A, and CXCL8 as hub nodes,could be used to effectively predict Covid-19 severity. Furthermore, the severity-associated SEMOs were found to be significantly correlated with CD4+ and monocyte cell proportions. These insights can be used to generate personalized nutraceutical regimes by ranking the relative severity risk associated with each SEMO. Thus, our pre-trained SEMO model can serve as a fundamental knowledge map when coupled with DNA methylation measurements, allowing us to simultaneously generate biomarkers, targets, re-purposing drugs, and nutraceutical interventions.