This paper introduces PerturbDiff, a novel computational framework designed to simulate how cells respond to biological disturbances like drugs or genetic modifications. Traditional models often struggle because single-cell sequencing is destructive, making it impossible to observe the same cell both before and after a treatment. PerturbDiff overcomes this by shifting the focus from individual cells to entire cell populations, treating these distributions as random variables in a high-dimensional mathematical space. By using a diffusion-based generative process and a distribution-matching objective, the model captures complex variability caused by hidden environmental factors. The researchers also employ a pretraining strategy using massive datasets to help the model generalize to new, unseen conditions even when data is scarce. Experiments show that this approach achieves state-of-the-art accuracy in predicting biological responses across various real-world benchmarks.

References:

• Yuan X, Liu X, Zhang Y S, et al. PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling[J]. arXiv preprint arXiv:2602.19685, 2026.