Cell-specific reconstruction improves gene regulatory network inference by resolving regulatory heterogeneity from single-cell multi-omics

Single-cell multi-omics provides an opportunity to link chromatin accessibility and gene expression within the same cell, yet most current gene regulatory network (GRN) inference methods derive regulatory weights through global modeling. Although effective for identifying consensus interactions, such strategies can average over cell-to-cell regulatory heterogeneity, obscuring context-specific edge activity and limiting the accuracy of downstream GRN prioritization. Here we present CellGRN, a plug-and-play framework that shifts GRN analysis from top-down global fitting to bottom-up cell-specific reconstruction. CellGRN takes candidate edges inferred by existing methods and reconstructs their activity in each cell by integrating transcription factor expression, chromatin accessibility, and target-gene expression. This reconstruction preserves regulatory heterogeneity at single-cell resolution, and thereby retains biologically meaningful signals that are diluted by direct global scoring. Across melanoma and PBMC datasets, benchmarking showed consistent improvements in sample-wise GRN inference over state-of-the-art backbone methods, including 3- to 10-fold improvements in cis-regulatory inference tasks. CellGRN also improves cell-type-specific regulatory inference and recovers dynamic regulon activity in spatial multi-omics despite severe expression dropout. Collectively, these results establish cell-specific reconstruction as a general strategy for gene regulatory network inference by resolving regulatory heterogeneity from single-cell multi-omics.