A single-cell velocity model you can perturb: it named three rescue targets for Arid1a-loss B cells, and mice confirmed all three

Single-cell multiome data can now yield a cell velocity field that can be knocked out, not just a time-ordered trajectory. MultiVelo also used RNA and chromatin accessibility to estimate velocities but gave no functional form for the underlying vector field, so trajectories could be drawn but not intervened on; Dynamo requires metabolic labeling, and scDiffEq uses RNA expression alone. On 7,166 wild-type mouse germinal center B cells, the learned field branched clearly into dark-zone centroblast cycling, the centroblast-to-centrocyte transition, and flow through recycling centrocytes back to the dark zone, while keeping exits toward prememory B cells and plasmablasts.

The model treats RNA expression and transcription-factor motif accessibility as two quantities that drive each other, solving their dynamics jointly in a low-dimensional latent space. Motif accessibility comes from chromVAR on scATAC as a proxy for transcription-factor regulatory activity; earlier RNA-based velocity models used spliced and unspliced counts only. Training supervises RNA velocities with cosine similarity to scVelo estimates for roughly 400 to 1,000 velocity genes, and adds a consistency loss so RNA and motif velocities trace similar paths in their respective latent spaces.

In silico perturbation of mutant cells can nominate targets that shift the loss-of-function phenotype back toward wild type. Arid1a-heterozygous germinal center B cells lose the centrocyte-to-dark-zone recycling trajectory, and the prememory B cell fraction rises from 5.42% in wild type to 20.09%. Perturbing candidate genes one at a time in the Arid1a-heterozygous model and ranking by the fraction of velocities shifted toward wild type put Ctcf, Stat3 and Bcl6 at the top; multiome profiling of Arid1a;Ctcf double-heterozygous mice showed the rapid prememory exit partially reversed, with a 64% RNA velocity rescue rate in centrocytes.

All three nominated targets were tested independently in living mice. An in silico nomination is not by itself functional evidence and needs genetic confirmation. Newly generated Arid1a;Bcl6 and Arid1a;Stat3 double-heterozygous mice, together with the existing Arid1a;Ctcf double-heterozygous mice, all restored the germinal center dark-zone to light-zone ratio to wild-type levels.

The report covers the mouse germinal center B cell setting: wild type, Arid1a-heterozygous, Ctcf-heterozygous, and the newly generated Arid1a;Ctcf, Arid1a;Bcl6 and Arid1a;Stat3 double-heterozygous models. It is aimed at researchers working on single-cell dynamics modeling and B cell lymphoma-related regulation. The source states no next experimental step.

The authors note that no single value of the regularization strength κ3 is best for all downstream tasks: higher values smooth latent times but may sacrifice accuracy of velocity changes after perturbation, and lower values do the reverse, so it must be chosen per task. Jacobian and in silico perturbation analyses give slightly different regulatory edges because the former uses infinitesimal and the latter finite perturbations. Known regulatory edges are reported as static in the literature, so dynamic comparison is only possible at the cell-state level.