ManifoldVelo fits cell order and RNA velocity together, cutting WT2 phase error 47% against recorded VeloCycle
Synopsis
A single-cell transcriptome snapshot carries both where a cell sits along a trajectory and which way its expression is moving, and these were usually estimated separately. ManifoldVelo places each cell on a differentiable unspliced/spliced RNA expression curve and matches the curve's derivatives to a transcription–splicing–degradation kinetic field, including between observed cells, so coordinates, curves and kinetics are fitted together. In multipeak cyclic simulations, spliced-only queries gave circular time errors of 0.64% and 7.40% with spliced-velocity cosines of 0.9776 and 0.9421; on WT2, phase error fell 47.0% relative to the recorded VeloCycle result, while VeloCycle did better on WT1 and the smooth cycle.
Interpretation
A cell's position in a developmental or cell-cycle progression and the direction its expression state is changing can now be estimated inside one continuous expression model. Earlier trajectory methods rebuilt order mainly from expression geometry, while RNA velocity methods gave direction from the unspliced-to-spliced relationship, and the two were typically fitted separately. In two multipeak cyclic simulations, on spliced-only query cells, ManifoldVelo gave circular time errors of 0.64% and 7.40% and spliced-velocity cosines of 0.9776 and 0.9421; on the same data VeloCycle gave time errors of 17.97% and 14.22% and velocity cosines of 0.7224 and 0.7985.
The fitted expression curve also serves as a reusable dynamic reference, so a new cell measured with spliced RNA alone receives a time coordinate and a velocity prediction. Reference parameters stay fixed during querying, so no decoder is refitted for a new sample. U2OS was evaluated by leaving out an entire plate and querying it with spliced RNA alone; the three-plate phase error was 11.481%, an 8.7% reduction from the recorded VeloCycle result of 12.575%.
Cycles and branches are written as two temporal-domain structures, letting one kinetic constraint cover closed loops and multiple paths from a shared root. A circular domain with a periodic decoder enforces closure for cyclic processes, while branching processes use several lineage curves that share one decoded root state, and an unbranched process uses a single open curve. In trifurcation and nested-branch simulations, ManifoldVelo's spliced-velocity cosines were 0.9735 and 0.9404 against 0.7624 and 0.7745 for TSvelo; time error was 1.75% versus 3.17% in trifurcation but 4.28% versus 3.33% in nested branches.
Perspective
The demonstrated scope covers six simulated trajectories, WT2 and WT1 mouse neurospheres, plate-held-out U2OS FUCCI cells, and four open-lineage datasets: pancreatic endocrinogenesis, dentate gyrus, LARRY and erythroid; real-data analyses use predefined roots and fixed configurations, while simulation anchors are inferred automatically from training unspliced/spliced RNA.
The authors report that VeloCycle performed better on WT1 and the smooth cyclic simulation, so the benefit depends on the dataset and trajectory representation; the inferred coordinate is normalized trajectory time rather than an independently calibrated duration in hours, simulation and full-S/U comparisons are mostly single archived runs, and the authors list donor-level validation and independent temporal calibration as future work.