Transfer Learning Prediction of Early Exposures and Genetic Risk Score on Adult Obesity in Two Minority Cohorts.
Journal:
Prevention science : the official journal of the Society for Prevention Research
PMID:
39913075
Abstract
Due to ethnic heterogeneity in genetic architecture, genetic risk score (GRS) constructed within the European population generally possesses poor portability in underrepresented non-European populations, but substantial genetic similarity exists across diverse ancestral groups. We here explore the prediction performance of early exposures and GRS on body mass index (BMI) through leveraging genetic similarity knowledge acquired from Europeans into non-Europeans. We present a linear mixed prediction model for BMI in three distinct UK Biobank cohorts under the transfer learning framework, where we consider Asians (n = 7487) and Africans (n = 7533) as target samples and Europeans (n = 280,575) as informative auxiliary samples. Besides environmental and behavior exposures, we incorporate multiple BMI-related variants, by which the GRS is constructed via transfer machine learning techniques informed by genetic similarity shared across target and auxiliary samples. The use of GRS gained more predictive odds for BMI than the model with traditional risk factors alone in the Asian and African cohorts, leading to an approximately 3.6% and 0.7% accuracy improvement in each target population. After borrowing genetic similarity from Europeans via transfer learning, the R increased to 0.270 for Asians and 0.302 for Africans, enhanced by 21.1% and 7.5%, respectively, compared to the early exposure-only models. We also provided evidence for the well-known conclusion that GRS constructed in the European population behaved poorly in non-Europeans. Prediction accuracy is greatly elevated in racial minority or underrepresented populations via the transfer learning method by leveraging shared genetic similarity from informative auxiliary populations.