Recommendation system-inspired learning of crossmodal interaction patterns between chromatin accessibility and gene expression
Year: 2026
Type: article
Abstract: Despite the growing availability of crossmodal single-cell sequencing, existing approaches are limited by underutilization of biological attributes, sparse modality signals, and insufficient modeling of regulatory interactions, limiting generalization in data-scarce settings and the discovery of peak-gene-level relationships. Interestingly, such challenges mirror those in recommendation systems, where side information, high sparsity, binary interactions, and cold start problems are extensively studied. Inspired by recommender models that infer user-item preferences from incomplete interaction graphs, we present CrossRec, which reformulates scATAC and scRNA data as an interaction graph to learn regulatory patterns across single-cell tasks. Its biologically-driven bipartite graph transformer learns interpretable attention over ATAC-RNA pairs. CrossRec outperforms existing methods across various tasks, remains robust under data-scarce settings, generalizes to unseen cell types and datasets, and enhances prior-based methods. Its learned regulatory maps provide cell- and cell-type-specific insights, resolving dynamic regulatory rewiring during cortical neurogenesis and perturbation-specific rewiring of drug-induced regulation in melanoma resistance.
Source: Nature Communications
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $6,990
Awards: G3795948297, G7689057940, G8195490573
OpenAlex ID: W7223456131