DeepVelo: deep learning extends RNA velocity to multi-lineage systems with cell-specific kinetics

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Current RNA velocity methods assume constant transcriptional rates and simple dynamics, which are often inaccurate for complex single-cell RNA sequencing (scRNA-seq) data.
• Biological systems exhibit heterogeneous cell populations with time-dependent kinetics and multiple differentiation lineages, posing challenges for existing RNA velocity estimation.

Purpose of the Study:

• To develop a generalized RNA velocity framework, DeepVelo, that accommodates time-dependent kinetics and multiple lineages in scRNA-seq data.
• To infer dynamic cellular processes including transcription, splicing, and degradation rates.
• To identify cell differentiation stages and key regulatory genes driving these processes.

Main Methods:

• Utilized a graph convolution network architecture for RNA velocity estimation.
• Developed a model capable of inferring time-varying rates of transcription, splicing, and degradation.
• Applied the method to analyze complex differentiation and lineage decisions in heterogeneous scRNA-seq datasets.

Main Results:

• DeepVelo successfully generalizes RNA velocity to complex cellular populations with time-dependent kinetics and multiple lineages.
• The method accurately infers dynamic cellular rates and determines individual cell stages within differentiation processes.
• Identification of functionally relevant driver genes regulating observed biological processes.

Conclusions:

• DeepVelo offers a robust approach to overcome limitations of existing RNA velocity methods in heterogeneous scRNA-seq data.
• The framework enables deeper insights into complex differentiation dynamics and lineage decisions.
• Demonstrates the utility of DeepVelo in studying both developmental and pathogenic biological processes.