UC San Diego

This dissertation explores machine learning applications in personalized genomics across three studies predicting cancer immunotherapy response, type 1 diabetes risk, and personalized cancer treatments.Chapter 1 investigated immune checkpoint blockade therapy response prediction using germline genetics and tumor-derived somatic features. Integrating germline immune-related variants with tumor somatic features significantly improved prediction accuracy over single feature approaches. The model identified non-linear interactions between genetic variants and revealed patient stratification based on MHC class I versus class II antigen presentation patterns, which correlated with distinct survival outcomes and checkpoint expression profiles. This demonstrated germline genetics' contribution to immune response and suggested targeting specific checkpoint ligands based on patient subgroups.Chapter 2 presented a large-scale genetic association and fine-mapping study of type 1 diabetes across diverse global populations including Latin American, African, and East Asian groups. A machine learning model called TIGRS outperformed existing genetic risk scores by capturing non-linear variant interactions, enabling highly accurate type 1 diabetes onset prediction. The model excelled particularly for individuals without high-risk HLA haplotypes and revealed extensive MHC and non-MHC loci interactions. Clustering analysis identified distinct genetic subtypes with varying onset ages and diabetic complication rates. This approach addresses genetic complexity that additive risk scores miss, providing an accessible prediction tool using common genetic variants while highlighting specific clinical complications requiring monitoring.Chapter 3 developed personalized cancer vaccine design methodology by training machine learning models on longitudinal tumor sequencing data. The innovation involved neoantigen elimination filtering to identify immunogenic tumor peptides and label integration across timepoints, generating individualized therapeutic targets. This precision medicine approach tailors cancer vaccines to each patient's unique tumor genetics, potentially improving treatment efficacy by focusing immune responses on patient-specific cancer antigens. Together, these studies demonstrate how integrating machine learning with genomic data enables more accurate disease prediction and personalized treatment strategies across multiple clinical applications.