Poster #14: Ben Stear MS

Pleiotropic Analysis of Neuroblastoma and Congenital Heart Defects using graph neural networks

Ben Stear, MS, Children's Hospital of Philadelphia ; Deanne Taylor, PhD, Children's Hospital of Philadelphia; Ahmet Sacan, PhD, Drexel University

Congenital heart defects (CHDs) and neuroblastoma (NBL) are two major pediatric disorders that, despite distinct clinical presentations, share a common developmental origin in neural crest cells. Emerging genetic and epidemiological evidence suggests that these diseases may arise through overlapping developmental mechanisms; however, conventional genomic analyses have limited ability to identify pleiotropic effects distributed across complex molecular networks. Here, we present an integrative graph-based framework that combines whole-genome sequencing data with a large-scale interactome to investigate shared genetic architecture between CHD and NBL. A heterogeneous graph was constructed linking patients, genetic variants, genes, and multiple layers of biological knowledge, including protein-protein interactions, transcription factor regulatory networks, pathway annotations, and microRNA-mediated regulation. Graph Attention Networks (GATs) were trained on this unified interactome to predict disease status while simultaneously generating interpretable attention coefficients that prioritize variants, genes, pathways, and regulatory relationships according to their contribution to model predictions. Multiple GAT architectures with increasing message-passing depth were evaluated to determine how progressively incorporating broader biological context influences disease prediction and feature prioritization. Model performance and biological relevance were further assessed through comparison with Random Forest classifiers and curated CHD- and NBL-associated gene sets. Rather than serving solely as a predictive framework, this approach leverages graph attention to identify candidate pleiotropic mechanisms linking congenital heart disease and neuroblastoma, providing a systems-level strategy for integrating rare genomic variation with multiscale biological networks to study complex developmental disease.