ADVANCING THE DISCOVERY AND IMPLEMENTATION OF PHARMACOGENOMICS USING GENOMIC BIOBANKS AND ARTIFICIAL INTELLIGENCE
Degree type
Graduate group
Discipline
Genetics and Genomics
Pharmacology, Toxicology and Environmental Health
Subject
Bioinformatics
Genomics
Large Language Models
Pharmacogenetics
Pharmacogenomics
Funder
Grant number
License
Copyright date
Distributor
Related resources
Author
Contributor
Abstract
While significant progress has been made in understanding the genetic underpinnings of drug response, clinical impact of these discoveries has been limited. Implementation of pharmacogenetics in routine care has been slow, and even when pharmacogenetic information is available for a patient, there are many un-accounted genetic and non-genetic factors which greatly influence drug response. To build evidence for the benefits of pharmacogenetic implementation, we used a large medical biobank to quantify rates of prescription for medications with known pharmacogenetic associations, as well as frequencies of pharmacogenetic variants. We demonstrated a high burden of pharmacogenetic variants, with many individuals receiving medication for which there were actionable pharmacogenetic guidelines based on their genetics. We also demonstrated that these burdens vary across ancestral populations, with certain populations disproportionately affected by specific variants. We then evaluated large language models (LLMs) as digital assistants for facilitating clinical pharmacogenetics. We built a comprehensive benchmark of pharmacogenetic questions and answers and used it to query several off-the-shelf models, discovering significant limitations that need to be accounted for before LLMs can be brought into the clinic. Lastly, we explored how summary statistics for external genome-wide association studies may be useful for building polygenic scores (PGS) for pharmacologic traits with low sample sizes. We used the AIDS Clinical Trials Group dataset to evaluate a multi-ancestry PGS for body mass index (BMI) which we repurposed towards predicting weight gain following antiretroviral therapy for HIV. While there was no association between genetic predisposition towards high BMI with increased weight gain, we successfully demonstrated the efficacy of our multi-ancestry BMI PGS in predicting BMI in a population of people living with HIV. Altogether, our results demonstrate avenues for improving pharmacogenetics both from the perspective of bringing existing knowledge into the clinic as well as generating new knowledge that transcends traditional monogenic pharmacogenetics.
Advisor
Kim, Dokyoon