Machine learning identifies SNPs predictive of advanced coronary artery calcium in ClinSeq(R) and Framingham Heart Study cohorts

One goal of personalized medicine is leveraging the emerging tools of data science to guide medical decision-making. Achieving this using disparate data sources is most daunting for polygenic traits and requires systems level approaches. To this end, we employed random forests (RF) and neural networks (NN) for predictive modeling of coronary artery calcification (CAC), which is an intermediate end-phenotype of coronary artery disease (CAD). Model inputs were derived from advanced cases in the ClinSeq(R) discovery cohort (n=16) and the FHS replication cohort (n=36) from 89th-99th CAC score percentile range, and age-matching controls (ClinSeq(R) n=16, FHS n=36) with no detectable CAC (all subjects were Caucasian males). These inputs included clinical variables (CLIN), genotypes of 57 SNPs associated with CAC in past GWAS (SNP Set-1), and an alternative set of 56 SNPs (SNP Set-2) ranked highest in terms of their nominal correlation with advanced CAC state in the discovery cohort. Predictive performance was assessed by computing the areas under receiver operating characteristics curves (AUC). Within the discovery cohort, RF models generated AUC values of 0.69 with CLIN, 0.72 with SNP Set-1, and 0.77 with their combination. In the replication cohort, SNP Set-1 was again more predictive (AUC=0.78) than CLIN (AUC=0.61), but also more predictive than the combination (AUC=0.75). In contrast, in both cohorts, SNP Set-2 generated enhanced predictive performance with or without CLIN (AUC> 0.8). Using the 21 SNPs of SNP Set-2 that produced optimal predictive performance in both cohorts, we developed NN models trained with ClinSeq(R) data and tested with FHS data and replicated the high predictive accuracy (AUC>0.8) with several topologies, thereby identifying several potential susceptibility loci for advanced CAD. Several CAD-related biological processes were found to be enriched in the network of genes constructed from these loci. In both cohorts, SNP Set-1 derived from past CAC GWAS yielded lower performance than SNP Set-2 derived from \"extreme\" CAC cases within the discovery cohort. Machine learning tools hold promise for surpassing the capacity of conventional GWAS-based approaches for creating predictive models utilizing the complex interactions between disease predictors intrinsic to the pathogenesis of polygenic disorders.