JuloVelo assigns each gene a kinetics pattern before training, resolving all five hippocampal neurogenesis branches
Synopsis
JuloVelo moves gene kinetics pattern identification ahead of neural network training: each gene is first classified as Circle, Complete induction, or Complete repression from its unspliced–spliced RNA relationship, and training starts from pattern-matched pretrained weights, steering cell-specific kinetics away from biologically implausible parameter regimes. Across eight biological datasets spanning erythrogenesis, neurogenesis, and endocrinogenesis, plus simulated data with known kinetics, it was more stable than constant-rate and cell-specific models, was the only method to resolve all five hippocampal neurogenesis branches, and used promoter chromatin accessibility to process multiome data.
Interpretation
RNA velocity inference gains a pre-training gene classification step: each gene is assigned to one of three kinetics patterns, Circle, Complete induction, or Complete repression, and the network then loads the pretrained weights matched to that pattern. Earlier cell-specific methods trained networks from unguided initialization, which can converge to kinetically inconsistent parameter estimates, particularly for genes with complex dynamic behavior. On simulated data with known ground-truth kinetics, JuloVelo recovered the expected velocity vector fields in all four regimes, Circle, Transcriptional boosting, Multiple induction, and Multiple repression, and its inferred transcription rates matched the true bimodal and boosting profiles.
Pattern-matched initialization keeps cell-specific transcription, splicing, and degradation rates within plausible ranges, so velocity directions follow the real differentiation path. Constant-rate models treat these three rates as constant or binary across cells and cannot represent a sharp mid-process surge in transcription rate. On mouse erythroid gastrulation, JuloVelo was the only method to infer a continuous forward flow from Blood progenitors 1 and 2 through Erythroid 1 and 2 to Erythroid 3, whereas scVelo produced disjointed and backward flows, cellDancer missed the progenitor transition, and DeepVelo reversed the vector field in mature erythroid cells.
In multiome data, promoter chromatin accessibility serves as an extra constraint on gene kinetics classification, resolving loci where RNA expression alone is ambiguous. Earlier methods did not use chromatin accessibility measured in the same cells to constrain transcriptional activity. On a mouse embryonic E18 brain multiome dataset, all four methods showed substantial errors in RNA-only mode, while adding promoter accessibility let JuloVelo restore a continuous flow connecting the upper and deeper cortical layers and raised both within-cell-type and across-boundary velocity stability.
Perspective
Validation covers eight biological datasets, including mouse erythroid gastrulation, hippocampal and pancreatic endocrinogenesis, skin and hair follicle, dentate gyrus neurogenesis, human bone marrow hematopoiesis, the mouse gastrulation E7.5 subset, and embryonic E18 brain multiome, plus simulated data with defined complex kinetics; the method takes unspliced and spliced RNA counts as input, and multiome mode additionally uses promoter chromatin accessibility.
Local, neighbor-based velocity inference can produce chaotic or undirected flow on isolated or highly discrete cell populations, and vector coherence in the dentate gyrus data improved noticeably only after isolated populations were removed; all splicing-based velocity methods are sensitive to gene annotation quality, as when anomalously high unspliced counts for Aff3 were attributed to the nested gene Aff3ir; in the dentate gyrus data, none of the four methods correctly inferred flow terminating in mature granule cells.