DiffDAG: Diffusion DAG Models for modeling Gene Perturbations

Understanding cellular responses to genetic perturbations in single-cell RNA-sequencing (scRNA-seq) is crucial for precision medicine, but challenges of heterogeneity, experimental costs, and data sparsity necessitates development of prediction tools. We introduce DiffDAG, a cell-type-agnostic framework that models regulatory interactions via a modular directed acyclic graph (DAG) in latent space to capture gene dependencies, and learns a conditional diffusion model using denoising bridges to transform unperturbed to perturbed expression distributions, enabling seamless generalization without cell-type-specific retraining. Extensive benchmarking on datasets involving SARS-CoV-2 infection, IFN$-\beta$ stimulation, systemic lupus, and cytokine-induced fate transitions shows DiffDAG surpassing baselines like scGen, with MMD reductions exceeding 95\% in CD4-T cells, alongside consistently highest Spearman correlations for gene expression ordering across all tested cell types, models and datasets.