Enhancing genotype-phenotype association with optimized machine learning and biological enrichment methods

Genome-wide association studies (GWAS) are surging again owing to newer high-quality T2T-CHM13 and human pangenome references. Conventional GWAS methods have several limitations, including high false negatives. Non-conventional machine learning-based methods are warranted for analyzing newly sequenced, albeit complex, genomic regions. We present a robust machine learning-based framework for feature selection and association analysis, incorporating functional enrichment analysis to avoid false negatives. We benchmarked four popular single nucleotide polymorphism (SNP) feature selection methods: least absolute shrinkage and selection operator, ridge regression, elastic-net, and mutual information. Furthermore, we evaluated four association methods: linear regression, random forest, support vector regression (SVR), and XGBoost. We assessed proposed framework on diverse datasets, including subsets of publicly available PennCATH datasets as well as imputed, rare-variants, and simulated datasets. Low-density lipoprotein (LDL) cholesterol level was used as a phenotype for illustration. Our analysis revealed elastic-net combined with SVR consistently outperformed other methods across various datasets. Functional annotation of top 100 SNPs from PennCATH-real dataset revealed their expression in LDL cholesterol-related tissues. Our analysis validated three previously known genes (APOB, TRAPPC9, and EEPD1) implicated in cholesterol-regulated pathways. Also, rare-variant dataset analysis confirmed 37 known genes associated with LDL cholesterol. We identified several important genes, including APOB (familial-hypercholesterolemia), PTK2B (Alzheimers disease), and PTPN12 (myocardial ischemia/reperfusion injuries) as potential drug targets for cholesterol-related diseases. Our comprehensive analyses highlight elastic-net combined with SVR for association analysis could overcome limitations of conventional GWAS approaches. Our framework effectively detects common and rare variants associated with complex traits, enhancing the understanding of complex diseases.