Identification of gene kinetics patterns improves RNA velocity and cell fate prediction using JuloVelo

JuloVelo improves RNA velocity model performance by initializing neural networks with gene kinetics patterns (Circle, Complete induction, Complete repression), outperforming constant-rate and cell-specific models across eight datasets.

• Why it matters: Existing RNA velocity methods assume constant or binary kinetics, which can miss heterogeneous splicing dynamics during differentiation and limit accurate future cell-state prediction.
• What they did: JuloVelo pre-determines per-gene kinetics patterns to set pattern-matched pre-trained weights, extends this to multiome using promoter chromatin accessibility, and adds in silico transcription factor perturbations.
• The result: Across eight datasets in erythrogenesis, neurogenesis, and endocrinogenesis, it consistently outperformed baselines, resolved a five-lineage hippocampal neurogenesis program, improved intrinsic/extrinsic stability in multiome, and reduced execution time by ~70% versus cellDancer.