Cisformer: a scalable cross-modality generation framework for decoding transcriptional regulation at single-cell resolution.

Single-cell multiomic technologies enable the joint analysis of different modalities, but face challenges due to experimental complexity. Current computational methods for single-cell cross-modality translation lack biological interpretability. Here, we present Cisformer, a cross-attention-based generative model tailored for cross-modality generation between gene expression and chromatin accessibility at single-cell resolution. Systematic benchmarking demonstrates the superior accuracy and generalization of Cisformer against existing methods. Cisformer leverages its inherent interpretability to precisely link cis-regulatory elements to target genes, facilitating the identification of functional transcription factors associated with tumorigenesis and aging. Overall, Cisformer is a powerful tool for single-cell multiomic data analysis.