Germinal centers are dynamic immune structures in which B cells repeatedly alternate between a proliferative dark zone and a selection-dominated light zone. During this cycle, centroblasts divide and diversify their immunoglobulin receptors, while selected centrocytes can receive survival signals, return to the dark zone and undergo further rounds of mutation and selection. Some cells eventually leave the reaction as prememory B cells or plasmablasts, which can give rise to memory B cells or plasma cells. Because single-cell sequencing destroys the cells it profiles, the sequence and regulation of these transitions must be inferred from a static snapshot, making it difficult to distinguish cell-fate drivers from correlations and to predict how a mutation will alter trajectories.

DynaVelo addresses that problem by treating RNA expression and transcription-factor motif accessibility as two connected dynamical systems. Variational encoders first map the two data types into lower-dimensional latent spaces, where a shared neural ordinary differential equation learns a vector field describing cellular change over an inferred latent time. Separate decoders translate that field back into gene-expression and motif-accessibility space, allowing the model to estimate RNA velocities for all modeled genes as well as velocities for transcription-factor motifs. Training combines reconstruction of the observed data, regularization of initial states and latent times, alignment with available scVelo RNA-velocity estimates for a subset of genes, and a consistency term requiring RNA and motif velocities to imply similar trajectories. The resulting model is generative, so it can create cells along inferred trajectories and simulate loss-of-function changes without requiring new cells for every perturbation.

Applied to 7,166 wild-type mouse germinal-center B cells, DynaVelo reconstructed the major branches of the reaction. Its model-derived vector field captured the centroblast cell cycle, the transition from centroblasts to light-zone centrocytes, the return of selected centrocytes to the dark zone, and exits toward prememory cells and plasmablasts. Its latent-time ordering more closely followed these biological transitions than the ordering produced by scVelo, which did not identify the bifurcation between dark-zone cycling and light-zone differentiation. A comparison with the multiome velocity method MultiVelo likewise found differences: MultiVelo did not recover the dominant centroblast-to-centrocyte direction and placed plasmablasts near the beginning rather than at a terminal exit. These results describe dynamics inferred from sampled cells and model outputs, not direct continuous observation of cells moving between zones.

Generated trajectories also exposed cell-state-specific regulatory patterns. Motif accessibility for Spi1 and Nfkb1 increased during the centroblast-to-centrocyte transition, whereas Spi1 remained stable and Nfkb1 decreased during dark-zone cycling. Along the route toward prememory cells, the model identified temporal changes in factors including Runx1 and Irf8, while the plasmablast branch showed accessibility and expression patterns involving Prdm1, Irf4 and Xbp1. The distinction between expression and motif accessibility was important: Prdm1 motifs were accessible in both prememory and plasmablast states, but Prdm1 RNA was expressed only in plasmablasts. The analysis therefore separated RNA-expression patterns from the accessibility of regulatory motifs, while the factors identified as possible drivers remain interpretations of temporal patterns rather than proof of causation in every inferred transition.

The same framework revealed how heterozygous deletions of Arid1a or Ctcf altered the inferred germinal-center landscape in samples containing 6,716 and 4,688 cells, respectively. Both mutants retained recognizable cell-cycle and differentiation branches but lacked the light-zone-to-dark-zone recycling path. Prememory cells accounted for 5.42% of wild-type cells, compared with 20.09% in Arid1aHet and 25.45% in CtcfHet germinal centers; the study also examined an Arid1a homozygous sample, in which the proportion was 44.96%. These are differences among genotypes, not measurements presented as a decline over time. The model associated the loss of recycling with altered activity of factors including Foxp1, Bach2, Spi1 and Ctcf, and with loss of a Foxo1-to-Cxcr4 regulatory relationship associated with migration back to the dark zone.

DynaVelo inferred dynamic gene-regulatory networks through two related calculations. Jacobian matrices measured the local derivative of one gene’s RNA velocity with respect to another gene’s expression, revealing how regulatory relationships varied among individual cells and along trajectories. Finite in silico knockouts instead set a transcription factor’s RNA and motif accessibility to their minimum observed values and calculated the resulting changes in velocity. The two approaches were not identical: Jacobians represented infinitesimal local changes, whereas simulated knockouts imposed finite changes across cells. Together, they recovered multiple previously reported relationships in wild-type cells while identifying edges that were missing or changed in the mutants. One example was the Spi1-to-Bcl-6 relationship, which disappeared along the centroblast-to-centrocyte-to-prememory trajectory in both heterozygotes. The sign of the modeled Bcl-6 influence on Bcl-2 also differed between wild-type and mutant cells.

To evaluate the model as a predictor of genetic loss, the researchers simulated Arid1a loss in wild-type cells and compared the predicted velocity changes with differential velocities inferred from Arid1aHet cells. Agreement was generally stronger for motif velocities than for RNA velocities across cell types. The simulation predicted reduced connectivity between centroblasts and centrocytes, loss of recycling, and a stronger route from centrocytes to prememory cells, matching the altered mutant structure. The study then perturbed candidate genes in the Arid1aHet model and searched for changes in RNA and motif velocities that pointed toward wild-type values. Ctcf, Stat3 and Bcl6 were among the strongest candidates, alongside Tox and Bcor. A velocity was counted as rescued when its direction of change matched the direction from mutant to wild-type in that cell state, making these computational predictions distinct from the subsequent in vivo outcomes.

Experimental profiling of Arid1aHet;CtcfHet double-heterozygous germinal centers showed a partial reversal of the rapid centrocyte-to-prememory exit characteristic of the Arid1a mutant. In centrocytes, the measured rescue rates were 38% for RNA expression, 64% for RNA velocity, 51% for motif accessibility and 46% for motif velocity. The defining Arid1aHet phenotype was an elevated dark-zone/light-zone ratio. Two newly generated compound-heterozygous models, Arid1aHet;Bcl6Het and Arid1aHet;Stat3Het, each restored that ratio to wild-type levels, providing direct functional support for two predictions. Together with the Ctcf double heterozygote, the experiments validated all three nominated targets. The results show that DynaVelo can nominate experimentally testable genetic interactions in this mouse system, but the study does not establish clinical effectiveness, and its predictions remain dependent on the quality of the training data, chosen genes, velocity guidance and model settings.

The experimental work used age- and sex-matched mice, with both males and females represented in each group, and the researchers reported no sex-dependent effects. Germinal-center B cells were profiled by paired single-cell RNA sequencing and ATAC sequencing, while flow cytometry measured cell-surface markers and the dark-zone/light-zone composition of germinal centers. The computational inputs included 2,000 highly variable genes, approximately 400–1,000 genes with scVelo velocity estimates, about 200 expressed transcription factors and roughly 20 genes of interest, with the union used for modeling. The study used separate wild-type, Arid1aHet, CtcfHet and compound-heterozygous samples, and it trained each model on a 90:10 training and validation split with early stopping. These design features support the model’s use as a trajectory and perturbation-analysis framework for this particular germinal-center system, while the experimental rescue results concern mouse phenotypes rather than human disease outcomes or therapeutic interventions.

Subject of Research: Computational modeling of mouse germinal-center B-cell dynamics and validation of predicted genetic rescue targets

Article Title: Deep dynamical models of single-cell multiomic velocities predict loss-of-function and rescue perturbations in B cells

Article References: Karbalayghareh, A., Pelzer, B., Chin, C. R., Melnick, A., Barisic, D., & Leslie, C. S. (2026). Deep dynamical models of single-cell multiomic velocities predict loss-of-function and rescue perturbations in B cells. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03348-1

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03348-1

Keywords: single-cell multiome, germinal-center B cells, neural differential equations, gene-regulatory networks, in silico perturbation, Arid1a, Ctcf, Bcl6, Stat3, epigenetic regulation, Deep, dynamical