Predicting the unseen: A diffusion-based debiasing framework for transcriptional response prediction at single-cell resolution

Predicting cellular responses to genetic perturbations is critical for advancing our understanding of gene regulation. While single-cell CRISPR perturbation assays such as Perturb-seq provide direct measurements of gene function, the scale of these experiments is limited by cost and feasibility. This motivates the development of computational approaches that can accurately infer responses to unmeasured perturbations from related experimental data. We introduce dbDiffusion, a generative framework that integrates diffusion models with classifier-free guidance derived from perturbation information, operating in latent space through a variational autoencoder. Diffusion models are probabilistic generative models that approximate data distributions by reversing a Markovian diffusion process, progressively denoising Gaussian noise into structured outputs. By exploiting biological similarities in gene expression profiles and relationships among perturbations, dbDiffusion enables the conditional generation of gene expressions for previously unobserved perturbations. In contrast to competing approaches, dbDiffusion does not rely on Large Language Model or foundation models, which have been found to yield unsatisfactory results. Rather, it leverages embeddings derived from measured perturbations to generalize to unseen perturbations, effectively transferring information across related experimental conditions. In benchmarking against state-of-the-art methods on Perturb-seq datasets, dbDiffusion demonstrates superior accuracy in predicting perturbation responses. A methodological innovation of dbDiffusion is the integration of prediction-powered inference, which corrects for biases inherent in generative models and enables statistically rigorous downstream tasks, including the identification of differentially expressed genes. By combining deep generative modeling with principled inference, dbDiffusion establishes a scalable computational framework for predicting and analyzing transcriptomic perturbation responses, significantly extending the utility of Perturb-seq experiments.