scLDM: a conditional diffusion framework for single-cell perturbation prediction.

scLDM accurately predicts single-cell responses to perturbations with over 90% accuracy across diverse biological datasets, outperforming existing methods.

• Why it matters: Understanding how individual cells respond to external stimuli is essential for uncovering gene regulation and advancing drug discovery, yet capturing the complex, non-linear relationships remains challenging.
• What they did: The authors developed scLDM, a generative model combining variational autoencoders and conditional diffusion processes, trained on six datasets covering various biological conditions to predict post-perturbation states.
• The result: scLDM not only achieves superior predictive performance but also offers high interpretability through perturbation embeddings aligned with known biological mechanisms, enabling reliable in silico perturbation screening.