https://www.nature.com/articles/s41556-025-01787-7
Resolving temporal dynamics at the cellular level is central to understanding embryonic development, tissue regeneration and disease progression. With the rapid growth of single-cell time-series atlases, international research teams have built large-scale reference atlases covering key processes such as mouse embryogenesis and early human embryo implantation. Existing computational methods, however, face serious challenges when applied across datasets: the discretization of time labels in conventional approaches struggles to accommodate varying sampling intervals, limiting their generalizability in fast-moving processes (such as embryo implantation) and in cross-species comparisons.
Dr. Yijun Liu from the team of Prof. Yi Chang at the School of Artificial Intelligence, Jilin University, published a study as first author in Nature Cell Biology, titled "TemporalVAE: atlas-assisted temporal mapping of time-series single-cell transcriptomes during embryogenesis." The study addresses a long-standing challenge of single-cell time-series atlas data in biomedical research: although international collaborations have produced large-scale single-cell time-series atlases such as those of mouse embryogenesis, providing important references for disease model construction, the complexity of cell states and the strong coupling between experimental batches and time labels still make temporal resolution within individual datasets difficult. The team innovatively proposed TemporalVAE, a deep generative model with a dual-objective formulation that can accurately infer the biological time of each cell through a compressed latent space, and even achieve zero-shot learning. Validation on mouse developmental atlases shows that the model can analyze million-scale cell datasets, supports accurate determination of developmental stages across platforms, and can identify temporally sensitive genes through virtual perturbation. In a study of human embryo implantation, TemporalVAE successfully identified cell developmental stages under both in vivo and in vitro conditions, and completed cross-species comparative analysis of human, cynomolgus macaque and marmoset embryos, demonstrating its broad application potential in the biomedical field.
The study also pioneered a virtual perturbation analysis strategy that quantitatively identifies temporally sensitive genes through gene expression intervention. This not only enhances the interpretability of the model, but also provides a new research approach for deciphering the mechanisms of cell fate determination. The successful completion of this work provides a powerful computational tool for developmental biology research and will drive innovation in disease model evaluation and regenerative medicine research. This paper was completed under the guidance of Prof. Yi Chang of the School of Artificial Intelligence, Jilin University, in collaboration with Research Fellow Dandan Cao of The University of Hong Kong–Shenzhen Hospital and Prof. Yuanhua Huang of The University of Hong Kong.
