Poster #3 - Fahimeh Palizban

Investigating pediatric cancer through deep learning-enabled genomics and phenomics

Fahimeh Palizban, Frank Mentch, Michael March, Xiang Wang, Deborah Watson, Joseph Glessner, Hakon Hakonarson.

Background: Pediatric cancers differ fundamentally from adult malignancies, often arising during development and characterized by rare germline variants, low mutational burden, and heterogeneous clinical outcomes. Conventional GWAS and PheWAS approaches struggle to capture pleiotropy, nonlinear interactions, and shared genetic architecture, limiting biomarker discovery in small pediatric cohorts. Methods: We developed a deep learning framework to integrate genomic and electronic health record (EHR) data of ~ 7000 samples (2000 cancer and 5000 control samples) from Center for Applied Genomics (CAG). The approach includes (1) a convolutional neural network with contrastive learning for multi-trait GWAS to disentangle shared versus trait-specific SNP effects; (2) a transformer-based PheWAS model to predict clinical phenotypes such as recurrence and survival, incorporating attention mechanisms for interpretability; and (3) translational pipelines mapping significant loci to pathways, stratifying patients into risk groups, and identifying therapeutic leads using DrugBank and Connectivity Map. Results: Preliminary genomic analyses highlight immune-related (HLA-C) and mucin family genes (MUC16, MUC19, MUC3A) as potential contributors to pediatric cancer biology. These loci are implicated in immune evasion, tumor microenvironment modulation, and disease progression. Our models demonstrate feasibility in capturing subtle, biologically meaningful variant-phenotype associations that are overlooked by conventional approaches. Conclusion: This work will introduce a scalable, interpretable AI-driven framework for pediatric cancer genomics and phenomics. By integrating multi-trait genomic data with clinical phenotypes, our approach supports risk prediction, patient stratification, and therapeutic discovery.