The article introduces Multi-resolution Variational Inference (MrVI), a sophisticated deep generative model designed for analyzing single-cell genomic data from large-scale studies. MrVI is created to overcome the limitations of current methods by enabling both exploratory analysis (stratifying samples into groups without prior cell-state definitions) and comparative analysis (evaluating cellular and molecular differences between predefined sample groups) at a single-cell resolution. Utilizing a hierarchical latent variable structure and counterfactual predictions, the model distinguishes between sample-specific effects and technical nuisance factors. The efficacy of MrVI is demonstrated through case studies involving cohorts with COVID-19 and inflammatory bowel disease (IBD), and a chemical perturbation screen, showing its ability to reveal clinically relevant and previously overlooked biological heterogeneity.

References:

• Boyeau P, Hong J, Gayoso A, et al. Deep generative modeling of sample-level heterogeneity in single-cell genomics[J]. Nature Methods, 2025: 1-11.