Advances and Future Prospects of Diffusion Models in Cellular Perturbation Modeling

Abstract

: Due to the fast development of single-cell transcriptomics technologies, society is provided unprecedented resolution for analyzing cellular heterogeneity and researching gene regulatory networks, and individuals believe that effective solutions to adequately reproduce the cellular perturbations should be urgently identified. Nonetheless, conventional computational algorithms can not efficiently describe non-linear behavior, transitions, or extrapolate across perturbations, so it becomes difficult to satisfy the practical requirements of precision medicine research. Recently, it has been shown that deep generative models (especially diffusion models) have tremendous benefits in cell perturbation prediction because of the excellent distribution learning and strong generation quality of these models. The provided paper is a systematic literature review of the underlying theoretical principles and technological advancements of diffusion models. It fully details all major contributions of representative models like Squidiff in predictions of responses to genetic perturbation and drug simulation. Moreover, it discusses the challenges of the field and its research opportunities depending on computational capabilities, biological insight, and the incorporation of multi-omics.