Advances in single-cell genomics have enabled the large-scale construction of organ atlases, providing unprecedented insights into cellular states and their perturbations, such as those induced by signaling, drugs, or diseases. By integrating multi-omic and spatial data, these atlases offer a rich foundation for applying machine learning to decipher cellular responses.

Generative AI is transforming scientific discovery by enabling researchers to explore uncharted territories, generate novel hypotheses, and simulate complex biological phenomena. In this context, we ask how generative AI can be leveraged to model single-cell variation, potentially paving the way for a single-cell foundation model.

Here, I will first introduce representation learning approaches for identifying the gene expression manifold and modeling perturbations, with a focus on drug or other perturbation-induced changes. I will then discuss recent extensions incorporating optimal transport and flow matching, and, if time permits, compare these to more prior-based modeling approaches.

Short Bio: Fabian Theis, holding PhDs in Physics and Computer Science, currently leads the Helmholtz Munich Computational Health Center and serves as the scientific director of both HelmholtzAI and Helmholtz Pioneer Campus HPC. As a Full Professor specializing in ‘Mathematical Modelling of Biological Systems’ at the Technical University of Munich, he also acts as Associate Faculty at the Wellcome Trust Sanger Institute in Hinxton, UK.