SCRIPT: Predicting Single-Cell Long-Range Cis-Regulation Based on Pretrained Graph Attention Networks.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Single-cell cis-regulatory relationships (CRRs) are crucial for understanding gene regulation and disease mechanisms.
• Current computational methods lack accuracy in predicting single-cell CRRs due to insufficient integration of biological principles and large-scale single-cell data.

Purpose of the Study:

• To develop a novel computational method, SCRIPT, for accurate inference of single-cell CRRs from transcriptomic and chromatin accessibility data.
• To improve the prediction of long-range CRRs and facilitate the identification of disease-causing variants.

Main Methods:

• SCRIPT utilizes graph causal attention networks, incorporating empirical CRR evidence.
• Representation learning is enhanced through pretraining on atlas-scale single-cell chromatin accessibility data.
• Validation involved cell-type-specific chromatin contact and CRISPR perturbation data.

Main Results:

• SCRIPT achieved a mean AUC of 0.89, significantly outperforming state-of-the-art methods (AUC: 0.7).
• SCRIPT demonstrated over a twofold improvement in predicting long-range CRRs (>100 Kb).
• Application to Alzheimer's disease and schizophrenia prioritized disease-causing variants and elucidated their cell-type-specific functional effects.

Conclusions:

• SCRIPT provides a robust framework for inferring single-cell CRRs, advancing genetic diagnosis and target discovery.
• The method uncovers molecular genetic mechanisms missed by existing computational approaches.
• SCRIPT offers a roadmap for understanding the functional impact of non-coding variants in disease.