Poster #48 - Shizhuo Mu

Deep Generative Modeling of Isoform Aware RNA Velocity in Long-read Single-cell RNA-seq

Shizhuo Mu1,2, Zhuoran Xu1,2, Joe Chan2, Haoran Zhang1, Quan Ma1, Qin Li1, Kai Wang1,2; 1University of Pennsylvania, Philadelphia, PA, 2Children's Hospital of Philadelphia, Philadelphia, PA

RNA velocity measures the temporal dynamics of gene expression by modeling the splicing dynamics from unspliced to spliced mRNA. However, current RNA velocity methods are developed for short-read data, capturing only gene-level transitions while overlooking isoform-specific changes within genes. Moreover, due to limited read lengths, short-read sequencing also struggles to distinguish unspliced from spliced transcripts, as reads mapping to shared exons may originate from any isoform, leading to inaccurate estimates of splicing dynamics. Multiple studies have reported that isoform proportions can change significantly between cell types even when overall gene expression remains constant, highlighting the need for isoform-level velocity analysis. With recent advancements in single-cell long-read sequencing, it is now possible to accurately quantify unspliced and spliced transcripts and evaluate RNA velocity at isoform resolution. We present IsoVelo, a deep generative model that infers isoform-resolved RNA velocity, overcoming the limitations of gene-level approaches with greater interpretability. IsoVelo learns RNA velocity kinetics on both gene and isoform levels, explicitly modeling the dynamic changes in isoform composition over time. In both simulation and real data analyses, IsoVelo uncovers isoform-level dynamic signals that are masked when using gene-level approaches. In summary, IsoVelo provides a more comprehensive view of cellular transitions by leveraging isoform-level resolution enabled by long-read single-cell RNA sequencing data.