Understanding how genes are turned on or off inside individual cells is a key challenge in modern biology. Researchers from Tongji University have developed a powerful new computational model called Cisformer to help solve this problem.

In single-cell biology, scientists often study different “modalities” or types of information from the same cell, such as RNA sequencing data to see which genes are active, and chromatin accessibility data to see which regions of DNA are open for gene expression. However, combining these datasets is extremely complex, and many existing computer models struggle to explain the underlying biology.

Cisformer uses an advanced AI architecture known as cross-attention to translate information between gene expression and chromatin accessibility. In simpler terms, it helps scientists understand how the structure of DNA in a cell influences which genes are turned on.

Overview of the Cisformer model

a The model architecture of Cisformer for single-cell RNA-to-ATAC (left) and ATAC-to-RNA (right) generation. Gene and peak features are encoded respectively, integrated through Transformer blocks with cross-attention mechanisms, and subsequently transformed into final outputs by an MLP. b The genes and chromatin peaks are selected or duplicated to generate gene-peak pairs as model input (top). Chromatin peak indices are encoded through a digit decomposition strategy, with 1,013,459 and 32,488 as representative examples (bottom). c The downstream applications of Cisformer, including validating biological interpretability (left), identifying cell-type-specific TFs in pan-cancer datasets (middle), and predicting aging-related CREs (right)

When tested against other methods, Cisformer outperformed them in both accuracy and its ability to generalize across different datasets. Importantly, it also offers biological interpretability, meaning it can reveal which DNA regions control specific genes. This helps identify key transcription factors, proteins that regulate gene activity and are often involved in processes like cancer development and aging.

By improving our ability to connect regulatory elements to their target genes at the single-cell level, Cisformer opens new opportunities to uncover how gene regulation changes across different cell types and disease states.

Availability – The source code and pre-trained models for Cisformer is available at GitHub (https://github.com/wanglabtongji/Cisformer) and Zenodo (https://doi.org/10.5281/zenodo.16991152)