SCRIPT

Most disease-associated genetic variants fall outside coding regions, and interpreting them requires knowing which regulatory element controls which gene — in which cell type. These cis-regulatory relationships (CRRs) are cell-type specific and frequently long-range, with enhancers acting across hundreds of kilobases. Correlation-based methods that pair chromatin accessibility with expression tend to degrade sharply as that distance grows.

SCRIPT (Single-cell Cis-regulatory Relationship Identifier based on Pre-Trained graph attention networks) infers single-cell CRRs jointly from transcriptomic and chromatin accessibility data. It was developed by Weidong Tian's group at Fudan University with the Shanghai Academy of Artificial Intelligence for Science and KAUST, first posted as a preprint in April 2025 and published in Advanced Science in August 2025.

The method combines two ideas. First, it follows the pretrain-then-apply pattern established by single-cell foundation models such as scGPT and Geneformer, but on chromatin rather than expression: a self-supervised graph autoencoder is pretrained on an atlas-scale scATAC-seq corpus to learn transferable representations of accessible regions. Second, it replaces generic attention with a graph causal attention network whose structure encodes empirical evidence about how regulatory elements act on transcription, rather than leaving that relationship entirely to be learned.

#Key Features

• Atlas-scale self-supervised pretraining: A masked graph autoencoder is pretrained on roughly 1.3 million human cells from a public scATAC-seq atlas (GEO accession GSE184462), giving downstream tissue-specific models a strong initialisation.
• Graph causal attention: The transcription-simulation module is built as a graph attention network constrained by empirical CRR evidence, so predicted regulatory links reflect a modelled mechanism rather than co-accessibility alone.
• Long-range regulation: Performance holds where correlation-based methods fall away, with a mean cell-level AUC of 0.90 for relationships spanning 100 kb to 1 Mb.
• Orthogonal validation: Predictions are benchmarked against cell-type-specific chromatin contact maps and CRISPR interference perturbation data, not only against co-expression.
• Open implementation: Code and usage documentation are released under the AGPL-3.0 licence.

#Technical Details

SCRIPT has two components. The self-supervised graph autoencoder (SSGAE) uses a two-layer graph attention network encoder and a single-layer decoder, with a node hidden dimension of 64, four attention heads, dropout of 0.5, and a masking ratio of 0.5; it is pretrained by reconstructing masked chromatin-accessibility graphs across the 1.3-million-cell atlas. The graph causal attention network (GCAT) then applies a single graph attention layer to simulate transcription from the learned representations, and is fit per tissue or downstream context.

On cell-type-specific chromatin contact benchmarks SCRIPT reaches a mean cell-level AUC of 0.89 and a mean reg-level AUC of 0.75, against 0.66/0.66 for LINGER and 0.70/0.61 for SCARLink, with mean AUPR ratios of 5.52 and 1.66. For long-range relationships beyond 100 kb it delivers an over twofold relative improvement in both cell-level and reg-level AUC over those baselines. Evaluation datasets span Hi-C contact maps, GTEx eQTLs, and paired scATAC/scRNA-seq cohorts.

#Applications

The practical target is non-coding variant interpretation. Applied to cortex data from Alzheimer's disease and schizophrenia cohorts, SCRIPT prioritises candidate causal variants and assigns them to target genes in a cell-type-resolved manner, producing mechanistic hypotheses that co-accessibility methods miss. The same output — a cell-resolved map of which element regulates which gene — is useful for target discovery, for interpreting GWAS loci in complex disease, and for designing follow-up CRISPR perturbation experiments.

#Impact

SCRIPT shows that atlas-scale self-supervised pretraining transfers usefully to regulatory-genomics tasks where labelled ground truth is scarce, and that the long-range regime is where that transfer pays off most. Its limitations are practical: the GCAT component is fit per tissue or downstream context rather than applied zero-shot, ground truth for CRRs remains indirect even when drawn from chromatin contact and CRISPRi data, and the pretraining corpus is human-only, so applying the model outside the tissues represented in the atlas is untested.