Estimating cell compositions and cell-type-specific expressions from GWAS data using invariant causal prediction, deep learning and regularized matrix completion

Integrating the cellular resolution of single-cell RNA sequencing (scRNA-seq) with the phenotypic depth of population-scale biobanks is essential for elucidating the cellular basis of complex diseases. However, this integration is often hindered by the limited sample sizes of scRNA-seq cohorts and the lack of cell-type resolution in massive biobank datasets. We present CausalCellInfer, a scalable computational framework designed to bring cellular resolution to bulk and genotype-imputed transcriptomes. CausalCellInfer utilizes an invariant causal prediction-inspired procedure (scI-GCM) to identify environment-stable marker genes, employs a parsimonious deep neural network for robust cell-fraction deconvolution, and leverages regularized matrix completion to reconstruct cell-type-specific (CTS) expression profiles. This architecture is specifically optimized for biobank-scale data, where technical heterogeneity and limited gene overlap are prevalent. Validated across simulated data, pseudo-bulk mixtures, and real PBMC datasets, CausalCellInfer demonstrated superior accuracy and computational efficiency compared to existing methods. Applied to [~]500,000 UK Biobank participants, the framework enabled cell-resolved analyses for 29 traits, identifying known pathological shifts, such as reduced pancreatic {beta}-cell proportions in Type 2 Diabetes, and uncovering novel biological signals, including disrupted excitatory neuron and oligodendrocyte interactions in depression. Furthermore, inferred CTS differential expression patterns showed significant concordance with independent single-cell studies and were enriched for OpenTargets disease genes. Overall, CausalCellInfer bridges the gap between single-cell insights and population-scale genomics, providing a powerful tool for systematic discovery of disease mechanisms at cellular resolution.