CausalCellInfer: Prioritizing Cell‐Type‐Specific Disease Associations From Genetically Predicted Transcriptomes in Biobank Cohorts

Integrating 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 is hindered by limited scRNA-seq sample size and the lack of cellular resolution in biobanks. We present CausalCellInfer, a scalable computational framework bringing cellular resolution to bulk and genotype-imputed transcriptomes. CausalCellInfer employs an invariant causal prediction-inspired method(scRNA-seq-based invariant generalized covariance measure, scI-GCM) to select environment-stable marker genes, a parsimonious deep neural network for cell-fraction deconvolution, and regularized matrix completion to reconstruct cell-type-specific (CTS) expression profiles. This design is optimized for biobank-scale data with technical heterogeneity and limited gene overlap. Validated across simulated data, pseudo-bulk and real peripheral blood mononuclear cell (PBMC) datasets, CausalCellInfer demonstrated superior accuracy and efficiency over existing methods. Applied to ∼500 000 UK Biobank (UKBB) participants across 29 traits, it captured known pathological shifts, such as reduced pancreatic β-cell proportions in type 2 diabetes, and identified novel biological signals, including disrupted excitatory neuron-oligodendrocyte interactions in depression. Inferred CTS differential expression patterns showed significant concordance with independent single-cell studies and were enriched for Open Targets disease genes. Overall, CausalCellInfer bridges single-cell insights and population-scale genomics, enabling systematic discovery of putative disease mechanisms at cellular resolution.