Using Topic Modeling via Non-negative Matrix Factorization to Identify Relationships between Genetic Variants and Disease Phenotypes: A Case Study of Lipoprotein(a) (LPA)

Abstract

Genome-wide and phenome-wide association studies are commonly used to identify important relationships between genetic variants and phenotypes. Most of these studies have treated diseases as independent variables and suffered from heavy multiple adjustment burdens due to the large number of genetic variants and disease phenotypes. In this study, we propose using topic modeling via non-negative matrix factorization (NMF) for identifying associations between disease phenotypes and genetic variants. Topic modeling is an unsupervised machine learning approach that can be used to learn the semantic patterns from electronic health record data. We chose rs10455872 in LPA as the predictor since it has been shown to be associated with increased risk of hyperlipidemia and cardiovascular diseases (CVD). Using data of 12,759 individuals from the biobank at Vanderbilt University Medical Center, we trained a topic model using NMF from 1,853 distinct phecodes extracted from the cohorts electronic health records and generated six topics. We quantified their associations with rs10455872 in LPA. Topics indicating CVD had positive correlations with rs10455872 (P < 0.001), replicating a previous finding. We also identified a negative correlation between LPA and a topic representing lung cancer (P < 0.001). Our results demonstrate the applicability of topic modeling in exploring the relationship between the genome and clinical diseases.\n\nAuthor summaryIdentifying the clinical associations of genetic variants remains crucial in understanding how the human genome modulates disease risk. Traditional phenome-wide association studies consider each disease phenotype as an independent variable, however, diseases often present as complex clusters of comorbid conditions. In this study, we propose using topic modeling to model electronic health record data as a mixture of topics (e.g., disease clusters or relevant comorbidities) and testing associations between topics and genetic variants. Our results demonstrated the feasibility of using topic modeling to replicate and discover novel associations between the human genome and clinical diseases.