Deciphering transcriptional regulation at single-cell resolution remains a major challenge. Existing GRN inference methods often depend on bulk-derived priors, which overlook cell-state differences, or on co-expression graph, which suffers from sparsity and limited interpretability. We present a scalable, end-to-end deep learning framework that integrates multi-omics preprocessing, graph construction, representation learning, and prediction in a unified workflow. The core model employs a multi-view attention mechanism that combines curated regulatory topology with condition-specific gene expression features, capturing both global structure and dynamic signals. Our framework reconstructs regulatory networks tailored to input conditions, avoiding biases from bulk priors. For example, it identified a bone marrow?specific BCL11A?DNMT1 interaction absent in cord blood predictions and prior networks, consistent with BCL11A’s repressive role via DNMT1-mediated hypermethylation. Altogether, this approach advances condition-specific, biologically grounded single-cell GRN inference with broad applicability to multi-omics integration and regulatory discovery.