Inferring differential dynamics from multi-lineage, multi-omic, and multi-sample single-cell data

Understanding the gene regulatory mechanisms that establish and maintain cell type identities is a central goal in cellular and developmental biology. Single-cell RNA sequencing and multi-omic profiling have revolutionized this field, enabling high-resolution investigation of gene expression dynamics across differentiation stages. RNA velocity, which estimates gene expression changes using mechanistic models, has emerged as a powerful approach for trajectory inference. Recent advances in RNA velocity methods address key limitations such as steady-state assumptions and lack of support for multi-omic data but still fall short in multi-sample integration and differential testing. To overcome these challenges, we introduce MultiVeloVAE, a probabilistic framework for multi-sample RNA velocity inference that integrates single-cell RNA and multi-omic data. MultiVeloVAE models gene expression on a shared time scale, accounts for lineage bifurcations, and enables statistical testing of velocity parameters. Our approach achieves a good balance between batch correction and biological variance conservation and outperforms existing methods in trajectory reconstruction. Using newly generated 10X Multiome datasets from human embryoid bodies and hematopoietic cells, we demonstrate that MultiVeloVAE provides novel insights into chromatin accessibility and gene expression dynamics during development. These results highlight the potential of MultiVeloVAE as a comprehensive tool for de novo multi-omic trajectory analysis and biological discovery.