Abstract 1677 Implementing a Support Vector Machine-based Computational Method to Prioritize Cardiovascular Disease-Risk Variants

Genome-wide association studies (GWAS) have linked multiple mutations (>500,000) to some type of genetic disease or trait. Over 90% of disease-associated mutations, such as single nucleotide polymorphisms (SNPs), have been mapped within the non-coding genome, which contain regulatory sequences bound by transcription factors (TF) to regulate gene expression. Disease-associated non-coding SNPs can alter gene regulation by disrupting TF binding sites (TFBS). However, due to the overwhelming and increasing number of disease-associated variants, prioritizing causal SNPs remains challenging. To address this challenge, we implemented support vector machine (SVM) and cardiac expression quantitative trait loci (eQTL) analysis to identify and prioritize potential cardiovascular disease (CVD)-causing SNPs for in vitro validation. Since gene expression is regulated by TFs in a tissue-specific manner, we aimed to determine if these SNPs could alter cardiac TF-DNA binding. We trained three large-scale gapped k-mer (LS-GKM) SVM predictive models using public ChIP-seq data from NKX2-5, TBX5, and GATA4, three cardiac TFs essential for proper heart development. The purpose of these LS-GKM SVM-based models was to identify binding sites for NKX2-5, TBX5, and GATA4. All models were successful in scoring TFBS with different affinities for NKX2-5 (AUROC = 0.955), TBX5 (AUROC = 0.921), and GATA4 (AUROC = 0.950) and were used to score all possible 2,097,152 non-redundant 11 bp oligomers (11-mers). The 11-mer scores were sorted and the 1,000 top-scoring sequences were used to generate a Position Weight Matrix (PWM). We observed similar motifs to those previously reported for NKX2-5, TBX5, and GATA4. We scored ∼520,000 fetal heart DNase I hypersensitivity genomic footprints (DGF) that occur within heart enhancers to identify genomic loci potentially bound by NKX2-5 or TBX5. To verify that our models could identify NKX2-5 and TBX5 binding sites and discern between DNA sequences with different affinities, we selected DNA sequences with the highest, middle, and lowest scores to test for in vitro binding through electromobility shift assays (EMSA). The results agreed with both model predictions, suggesting they can be used to predict the effect of SNPs on TF-DNA binding by scoring reference and mutated sequences. To test our models, we identified CVD-associated SNPs from the GWAS catalog and intersected them with human enhancers active in the heart and DGF from cardiac tissue. CVD-associated SNPs occurring within human heart enhancers and DGF were expanded to include variants in linkage disequilibrium (LD). CVD-associated SNPs and variants in LD were further filtered through the Genome Tissue Expression (GTEx) database to identify SNPs with genotype-dependent expression in cardiac tissue. To validate our computational method, we selected two SNPs that were predicted to disrupt NKX2-5- (rs59310144) and TBX5- (rs7612445) DNA binding and are associated with differentially expressed genes (RNASEH2B and GNB4, respectively) in cardiac tissue. Both variants altered NKX2-5- and TBX5-DNA binding activity as determined through EMSA. Our findings suggest that regulation of RNASEH2B and GNB4 can be further studied to establish rs59310144 and rs7612445 as causal CVD-risk variants. In short, this approach can provide crucial information to implement during in vivo experiments or clinical research to address genetic diseases caused by non-coding SNPs. We believe this integrative approach, which considers tissue-specific eQTL analysis and SVM-based TFBS classification, is a scalable method that can be applied to multiple types of human diseases. This project was supported by NIH-SC1GM127231, NSF [1736026], INBRE [P20GM103475], NIH RISE Fellowship (5R25GM061151-20), NSF [IQ BIOREU 1852259], NSF BioXFEL (STC1231306), NSF HRD-2008186. NSF REU: PR-CLIMB (2050493), NIH 1T34GM145404-01A1, and NIH ID-GENE (1R25HG012702-01).