New Model Predicts Single-Cell Perturbation Responses

Researchers have developed CisTransCell, a new multi-modal framework aimed at improving predictions of single-cell transcriptional responses to genetic perturbations. This model is particularly effective in zero-shot settings, where it can predict the effects of gene combinations not encountered during its training phase. The challenge in predicting perturbation effects lies in their dependence on intricate biological factors beyond simple gene expression. These include how gene products influence other genes and proteins, how downstream factors interact with cis-regulatory elements, and the specific regulatory programs active within a cell. CisTransCell addresses this complexity by augmenting each gene with two crucial prior information sources: a regulatory-sequence prior detailing how the gene is controlled, and a coding-sequence prior explaining the gene product's function. By combining these priors with the cell's expression state, CisTransCell models perturbation response as a cascade from gene function through regulatory control to observable transcriptional changes. Experimental results on benchmark single-cell perturbation datasets demonstrate that CisTransCell achieves strong performance in zero-shot prediction tasks, indicating its potential to advance understanding in single-cell biology.