SCRIPT: Predicting Single-cell Long-range Cis-regulation Based on Pretrained Graph Attention Networks

Single-cell cis-regulatory relationships (CRRs) are essential for deciphering transcriptional regulation and understanding the pathogenic mechanisms of disease-associated non-coding variants. Existing computational methods struggle to accurately predict single-cell CRRs due to inadequately integrating causal biological principles and large-scale single-cell data. Here, we present SCRIPT (Single-cell Cis-regulatory Relationship Identifier based on Pre-Trained graph attention networks) that infers single-cell CRRs from transcriptomic and chromatin accessibility data. SCRIPT incorporates two key innovations: graph causal attention networks supported by empirical CRR evidence, and representation learning enhanced through pretraining on atlas-scale single-cell chromatin accessibility data. Validation using cell-type-specific chromatin contact data demonstrates that SCRIPT achieves a mean AUC of 0.9, significantly outperforming state-of-the-art methods (AUC: 0.68). Notably, SCRIPT obtains a threefold improvement in predicting long-range CRRs (>100 Kb) compared to existing methods. Applying SCRIPT to Alzheimers disease and schizophrenia, we establish a framework for prioritizing disease-causing variants and elucidating their functional effects in a cell-type-specific manner. By uncovering molecular genetic mechanisms undetected by existing computational methods, SCRIPT provides a roadmap for advancing genetic diagnosis and target discovery.