Exploration of candidate genes associated with rare SNVs in pulmonary stenosis using whole-exome sequencing and machine learning.

Pulmonary stenosis (PS) is a common form of congenital heart disease (CHD) that impairs cardiopulmonary function and can be life-threatening in severe cases. As a complex polygenic disorder, the genetic basis of PS remains incompletely understood. Rare pathogenic single nucleotide variants (SNVs) were identified from whole-exome sequencing (WES) data of 185 sporadic PS patients and 100 healthy controls using multiple pathogenicity-filtering strategies. Gene-level burden test was performed, with complementary analysis using sequence kernel association test-optimal (SKAT-O). Three machine learning algorithms-least absolute shrinkage and selection operator (LASSO), random forest (RF), and extreme gradient boosting (XGBoost)-were applied to prioritize candidate genes. The overlap between machine learning-based selections and burden test results was systematically evaluated. Final candidate genes were further prioritized through protein-protein interaction (PPI) network analysis, and their expression in human pulmonary artery endothelial cells (HPAECs) was assessed by reverse transcription quantitative polymerase chain reaction (RT-qPCR). Comparative analyses showed that different machine learning algorithms exhibited distinct feature selection patterns, with RF demonstrating the highest concordance with burden test results. A total of 17 candidate genes were prioritized (HAND2, SETD2, KDM6B, NCOR2, FLNB, NOTCH3, DNAH5, PLEC, COL5A1, KIF7, CLTCL1, XRN1, ITPR2, SCRIB, PYGB, IQGAP3, and SHC2). These findings indicate that machine learning can complement conventional gene-based analyses of WES data. This study provides a set of candidate genes associated with PS and offers a basis for further investigation of its genetic architecture.