Abstract 041: A Novel Machine Learning Strategy to Integrate Multi-Omics Data and Detect Genomic Loci and Gene-Environment Interactions for LDL Cholesterol

Introduction: LDL cholesterol (LDLC) is highly heritable, and understanding its mechanism is crucial for cardiovascular disease treatment and prevention. While multi-omics data is widely available, integrating different omics in etiologic research of LDLC is limited due to computational challenge. Machine learning (ML) methods can address the challenge but are primarily used for risk prediction, and their “black box” nature complicates model interpretation. Hypotheses: We hypothesized that top important single nucleotide polymorphisms (SNPs) ranked by an ML-based genomic model contained novel genetic loci for LDLC, and that adding metabolomics data to the genomic model would change SNP importance due to mediation by internal metabolites and interaction with external metabolites. Methods: We developed a random-forest based genomic prediction model for LDLC using genome-wide SNPs among 1239 participants of the Bogalusa Heart Study (BHS), then incorporated 792 known metabolites to develop a multi-omics model. We compared variable importance matrix (VIMs) for SNPs between models. For SNPs losing importance, we tested whether metabolites mediated their associations with LDLC; for those gaining importance, we tested their interactions with all metabolites on LDLC. All analyses controlled for age, sex, race, and the first 10 genetic principal components. Results: VIMs changed significantly between the genomic and multi-omics models. Among the 20 SNPs with the largest drop in VIMs, we found that ( Table 1 ): 1) 256 metabolites, mostly (n=252) in lipid pathways, mediated ( P <0.05) their associations with LDLC. 2) These metabolites were top-ranked inputs in the multi-omics model. 3) Twelve loci have been reported by GWAS of lipid phenotypes, including 5 for LDLC. 4) Five genes, ELOVL6 , VAV2 , CDH10 , ACKR3 , and LMF1 were not reported by lipid GWAS, but were involved in lipid metabolism in animal models. For the 20 SNPs with the largest increase in VIMs, interacting metabolites were of non-lipid pathways or xenobiotics ( Table 2 ). Conclusion: We developed a novel ML strategy to integrates genomics and metabolomics data and provided empirical evidence of using it to discover new loci and gene-environment interactions for LDLC. This strategy can be expanded to other omics.