FocusMR: An Attention-Based Single-Cell Mendelian Randomization Framework to Map Cellular Contexts at Candidate Genes

Single-cell cohorts with matched genetic data offer new opportunities to uncover the cellular contexts through which genetic variants influence complex traits and disease.However, existing causal inference methods rely on cell-type–level pseudobulks and fail to resolve relevant cell states at higher resolution.We introduce FocusMR, a framework that integrates two-sample Mendelian Randomization (MR) with attention-based multiple instance learning (MIL) to localize the specific cell states mediating genetic effects.FocusMR learns attention weights over cells based on genotype–expression consistency at MR-supported loci, generating state-aware pseudobulks that can be transferred to independent individuals for validation and downstream analysis.Applied to eczema in immune single-cell data from OneK1K, FocusMR identifies interpretable sub-cell-type interactions and highlights disease-relevant cellular heterogeneity.By combining causal inference with interpretable cell-state modeling, FocusMR provides a principled approach to map genetic risk to specific cellular contexts, supporting early-stage target discovery.