Development of Statistical Methods for Multi-ancestry Transcriptome-wide Association Studies

Although Genome-wide association studies (GWASs) have successfully identified many risk variants associated with complex human phenotypes, most variants are located in the non-coding region, suggesting their regulatory role through the change of gene expression. By integrating GWASs and resources from expression quantitative trait loci (eQTL) mapping studies, transcriptome-wide association studies (TWASs) seek to determine the influence of genetically regulated gene expression (GREX) on complex traits. However, the diverse genetic backgrounds in eQTL samples presents significant challenges for conventional TWAS, which often rely on single-ancestry eQTL models. These methods are usually not generalizable to other ancestries due to variations in genetic architectures across different ancestries. Besides, the polygenic architecture of complex traits remains a major issue that causes inflation in TWAS. To address above challenges, we propose a unified statistical framework for multiancestry TWAS. In our pilot study, we propose a TWAS method to integrate multiancestry eQTL data and account for the polygenicity of complex traits (MAPO). A distinctive feature of MAPO is its ability to leverage ancestry-consistent eQTL signals, which enables more accurate characterization of eQTL effects across diverse ancestries, yielding enhanced TWAS power in a target ancestry. MAPO also explicitly characterizes the polygenic effects through its model design, producing well-controlled type-I errors. Unlike existing two-step TWAS procedures, MAPO employs a joint likelihood model that accounts for prediction uncertainty, which ensures statistical accuracy when working with eQTL data from under-represented ancestries. Preliminary results demonstrate MAPO's enhanced power and controlled type-I error compared to existing methods. Notably, we identified novel gene associations linked to complex traits, such as asthma, while successfully replicating known associations. Given the promising results, we propose several extensions for MAPO. First, it will be generalized to work with summary-level data, broadening its applicability to various phenotypes. Second, we plan to enhance computational efficiency by introducing a Monte-Carlo Expectation-Maximization algorithm. Third, we aim to extend MAPO to infer ancestry-specific gene-trait associations, facilitating the investigation of crossancestry heterogeneity in TWAS effects. The novelty of this research lies in its innovative unified model that utilizes multiancestry eQTL resources to overcome the generalizability issue of traditional TWAS methods. The unified model bridges the gap between diverse genetic backgrounds while naturally accounts for the uncertainty that was ignored by conventional methods. We believe that the proposed framework will advance the field of human genetics and deepen our understanding of complex traits across global populations, ultimately offering equitable healthcare and therapeutic guidance.