DeepCNet: A multimodal deep learning model for predicting cell type-specific gene expression and promoter-enhancer...

Understanding promoter-enhancer (PE) interactions is essential for elucidating gene regulation mechanisms. Although existing computational methods leverage DNA sequence to predict gene expression, many were developed primarily for bulk-level genomic data and emphasize predictive accuracy rather than regulatory inference. Here, we propose DeepCNet, a multimodal deep learning framework that predicts cell type-specific gene expression and infers cell type-specific PE interactions, enabling the construction of transcription factor gene regulatory networks (TF-GRNs) from single-cell multiome data. Evaluated on two single-cell multiome datasets from blood and brain tissues, DeepCNet consistently outperforms state-of-the-art methods in predicting cell type-specific gene expression. Importantly, PE interactions inferred by DeepCNet showed greater concordance with orthogonal chromatin interaction and CRISPR perturbation data than competing approaches. Analysis of cell type-specific GRNs identified TF hubs and highlighted both cell type-specific and shared regulatory programs. DeepCNet further extends to two-condition single-cell multiome studies, enabling the identification of differential PE interactions and characterization of TF-GRN alterations across conditions. As an illustrative application, we applied DeepCNet to single-cell multiome data from one Alzheimer's disease (AD) and one healthy control sample, identifying candidate differential PE interactions at AD risk loci and candidate TF hubs associated with AD-related pathways. Together, these results demonstrate the potential of DeepCNet to characterize transcriptional regulatory programs and cooperative TF interactions underlying cell type- and condition-specific gene expression.