Estimating cell compositions and cell-type-specific expressions from GWAS data using invariant causal prediction, deep learning and regularized matrix completion: Bridging GWAS and single-cell resolution in Biobank-scale studies

Abstract Dissecting large bulk RNA-seq data into cell-type proportions and cell-type-specific expression profiles has the potential to significantly enhance our understanding of disease mechanisms at the cellular level. While single-cell RNA sequencing provides detailed cellular insights, its application is limited by small sample sizes and cost constraints. Conversely, large-scale GWAS datasets offer extensive sample sizes but lack cell-type resolution. We present CausalCellInfer, a framework that integrates invariant causal prediction, deep learning and regularized matrix completion to identify critical cell markers, deconvolute cell proportions and estimate cell-type-specific(CTS) expression profiles. We pioneered the application of the proposed framework to imputed expression data from large-scale genome-wide association studies (GWAS), enabling cell-type level analysis in biobank-scale datasets. We validated CausalCellInfer against state-of-the-arts methods like CIBERSORTx, DWLS, Scaden, and MuSic using real and pseudo-bulk samples. Our framework consistently outperformed others with significantly higher concordance correlation coefficient (CCC), lower mean absolute error (MAE) and root mean square error (RMSE) while demonstrating superior computational efficiency. Application to the UK Biobank revealed novel biological insights across 24 phenotypes. We deconvoluted tissue-specific cell proportions and estimated CTS profiles across the phenotypes. We revealed that cell-type proportions were associated with disease susceptibility, such as decreased alpha and beta cells in T2DM patients. Furthermore, CausalCellInfer achieved high positive predictive values in identifying cell-type-specific differentially expressed genes. Overall, CausalCellInfer represents a significant advancement in integrating single-cell resolution with biobank-scale data and comprehensive clinical phenotypes, providing a powerful tool for elucidating disease mechanisms at the cellular level.