RNA velocity inference based on graph transformer

RNA velocity analysis is a crucial method for interpreting dynamic changes in cell states from single-cell sequencing data. Despite advances in modeling approaches based on ordinary differential equations and neural networks, challenges remain in capturing complex transcriptional dynamics and long-range dependencies between cells. We present GTVelo, a graph transformer-based neural ordinary differential equation model. By incorporating a multi-head attention mechanism and a multi-origin state mechanism, GTVelo mitigates the limitations of models relying on a single initial state. It enables capture of potential long-range cellular associations beyond local neighborhood graphs, flexible inference of multi-branch trajectories, and integration of multi-omics data. Across multiple benchmark datasets, GTVelo achieves competitive performance and outperforms compared methods on standard evaluation metrics in most cases. GTVelo provides a reliable computational tool for investigating complex cellular dynamics, with broad application potential in single-cell trajectory inference and multi-omic data analysis.