Large Models for Single-Cell Omics and Drug Discovery: Data, Pretraining, and Closed-Loop Environment

This presentation will discuss recent advances in foundation models for single-cell omics and therapeutic discovery, focusing on the development of scGPT, a generative transformer model trained on over 33 million single-cell profiles. It will cover key design principles—such as generative pretraining and multi-task alignment—that enable broad applications including cell type annotation, perturbation response prediction, and reference mapping. The talk will also highlight emerging multimodal extensions such as scGPT-spatial and MethylGPT, and emphasize the importance of perturbational training for modeling cellular dynamics. Finally, it will introduce LUMI-lab, a closed-loop self-driving platform developed by Dr. Cui and collaborators, which integrates model-guided design, synthesis, and validation to accelerate mRNA delivery—demonstrating a scalable framework for virtual cell modeling and active learning in biology.