Abstract
Transcriptome-wide association studies (TWAS) have successfully identified genes associated with complex traits and diseases, but most rely on bulk transcriptome data, overlooking cell-type-specific contexts. Population-scale single-cell RNA sequencing data now enable such analyses, but present unique challenges due to strong noises, technical variations, and high sparsity. Here, we propose scTWAS, a statistical method to conduct cell-type-specific TWAS using single-cell data. Leveraging a latent-variable model and moment-based estimation to address the challenges of single-cell data, scTWAS consistently improves the prediction of genetically regulated gene expression across cell types in both blood and brain tissues. Compared to existing methods, scTWAS identified substantially more gene-trait associations across 29 hematological traits and three immune-related diseases in immune cell types. An application to Alzheimer's disease also revealed cell-subtype-specific associations, including MS4A6A in disease-associated microglia and PPP1R37 in both inflammatory microglial and astrocyte subtypes.