2026-10-23
A latent neural ODE names Ctcf, Bcl6 and Stat3 as Arid1a rescues in germinal-center B cells; all three double-het mouse lines restore the dark-zone ratio.
A single-cell multiome profile is one snapshot. RNA and chromatin accessibility are measured in the same cell, and the cell is already lysed. Fate has to be inferred from a cloud of cells. RNA velocity, estimated from spliced and unspliced reads, covers only a subset of genes and does not hand back a vector field you can intervene on. MultiVelo adds a gene-level chromatin ODE and can draw trajectories, but still cannot be perturbed. Dynamo can simulate perturbations, yet it wants metabolic labeling, which many in vivo systems cannot provide. scDiffEq models RNA alone.
Mouse germinal-center B cells sit in that gap. Centroblasts proliferate in the dark zone, become centrocytes in the light zone, then some recycle to the dark zone for somatic hypermutation and affinity maturation. Others leave as pre-memory B cells or plasmablasts. Heterozygous loss of Arid1a, a BAF subunit often altered in human B-cell lymphoma, or of the insulator Ctcf, breaks this circuit. What was missing is a model that can knock a gene out in silico and name which transcription factor to hit next.
DynaVelo, from Christina Leslie's group at Memorial Sloan Kettering and collaborators at Weill Cornell, is a latent neural ODE on paired scRNA-seq and scATAC-seq. chromVAR turns accessible peaks into transcription-factor motif activity, about 200 motifs after filtering. The RNA side keeps highly variable genes, genes with an scVelo velocity, expressed transcription factors, and a short list of germinal-center genes, roughly 3,000 to 4,000 features in the union.
Each modality goes through its own shallow variational encoder into a 50-dimensional latent space. A neural ODE with a 200-unit hidden layer, integrated by dopri5, evolves the joint latent state. Decoders map it back to expression and motif accessibility. Observed velocities are the decoder Jacobian times the latent velocity, so velocity is the time derivative of the reconstructed state. Latent time has a Beta(2, 2) prior. The initial latent state has a standard normal prior.
The loss has four terms. Reconstruction is a Gaussian log-likelihood with variance fixed at 0.01. A cosine term matches predicted RNA velocity to scVelo on the 400 to 1,000 genes where scVelo has an estimate. A consistency term pushes the cell-by-cell cosine matrices of RNA velocity and motif velocity toward each other. KL terms keep the initial state and latent time near their priors. On this dataset the velocity and consistency weights are 10,000. The prior weight is 1,000 when the goal is a smooth latent time, and 1 when the goal is perturbation accuracy. The authors treat that weight as task-dependent.
With a functional vector field, regulation is read two ways. Jacobians are infinitesimal: how a tiny change in one gene's expression moves every other RNA or motif velocity, evaluated per cell. In silico knockout is finite. A transcription factor has both its RNA and its motif accessibility set to the minimum observed across cells. Other genes have only RNA set to the minimum. A feature counts as rescued in the mutant model when that knockout shifts its velocity toward the wild-type value. The rescue rate is the fraction of significant features whose sign agrees, averaged over cell types.
On 7,166 wild-type germinal-center B cells, the learned field separates dark-zone cell cycle, the centroblast-to-centrocyte transition, exits toward pre-memory cells and plasmablasts, and recycling from the light zone back to the dark zone. Latent time places pre-memory cells and plasmablasts at the end and a subset of centroblasts at the start. scVelo's latent time misses the dark-zone versus light-zone split, and its motif-space ordering is not usable. MultiVelo draws the dominant flow from centrocyte to centroblast, opposite the known direction, and treats plasmablasts as a start state.
Along generated trajectories, Spi1 and Nfkb1 motif accessibility rise monotonically from centroblast to centrocyte and do not do so across dark-zone cell cycle. That matches experiments in which Spi1 acts as a pioneer factor and then recruits NF-κB. Motif dynamics are clearer than the RNA of those two factors.
Arid1a heterozygous cells (6,716) and Ctcf heterozygous cells (4,688) both lose light-zone-to-dark-zone recycling. The pre-memory fraction rises from 5.42% in wild type to 20.09% in Arid1a heterozygotes, 25.45% in Ctcf heterozygotes, and 44.96% in Arid1a homozygotes. The wild-type edge in which Spi1 upregulates Bcl6 disappears along the centroblast-to-pre-memory path in both mutants. Foxo1 upregulation of Cxcr4, used when centrocytes migrate back to the dark zone, is lost in Arid1a heterozygous centrocytes.
An in silico Arid1a knockout in the wild-type model was compared with real heterozygotes. Overlap of significant velocity changes is a Jaccard index. Direction agreement is the fraction of that intersection with the same sign, at adjusted P < 0.05. Across 11 cell types, motif-velocity overlap is 0.93 to 0.99 and direction agreement is 0.75 to 0.88. RNA-velocity overlap is 0.49 to 0.76 and direction agreement is only 0.17 to 0.52. Chromatin-side velocity shifts are predicted more reliably than transcript-side shifts.
Perturbing transcription factors inside the Arid1a heterozygous model ranks Ctcf, Stat3 and Bcl6 among the stronger rescue candidates, with Tox and Bcor in the same region of the plot. The Ctcf prediction lines up with independent work on SWI/SNF, CTCF and cohesin. New multiome profiles of Arid1a/Ctcf double heterozygotes show a partial reversal of the rush from centrocytes toward pre-memory. In centrocytes, the cell type hit hardest in Arid1a heterozygotes, the fraction of features moved back toward wild type is 38% for RNA expression, 64% for RNA velocity, 51% for motif accessibility and 46% for motif velocity. RNA-velocity rescue across cell types sits around 0.63 to 0.71. Expression rescue ranges from 0.03 to 0.50.
Bcl6 and Stat3 were tested in new mice. Arid1a/Bcl6 and Arid1a/Stat3 double heterozygotes both restore the dark-zone to light-zone ratio to wild-type levels by flow cytometry. The Bcl6 cohort has 6 mice per genotype. The Stat3 cohort has 10 wild type, 10 Arid1a heterozygous, 6 Stat3 heterozygous and 4 double heterozygous mice, 54 animals in total, compared with a two-sided Welch t-test. All three nominated targets rescue the same in vivo phenotype.
Few single-cell dynamical models close the loop from an in silico knockout to newly built double-heterozygous mice, and have all three nominated targets hit. For germinal-center epigenetics and B-cell lymphoma, reducing Ctcf, Bcl6 or Stat3 on an Arid1a heterozygous background is a genetic interaction worth following. The Ctcf arm also has multiome data, and the predicted velocity shift matches the double mutant.
The method wants paired RNA and ATAC, and it is trained separately per genotype. Five to twenty thousand cells take up to about three hours on one A6000. This is not a drop-in rescue-target caller for arbitrary tissues. Against scVelo and MultiVelo, the gain is a branching latent time plus a vector field you can perturb. The evidence is still one cellular system.
RNA-velocity direction agreement tops out at 0.52 and bottoms at 0.17, against 0.75 or better on the motif side. Claims about whether a gene's transcriptional velocity rises or falls rest on the weaker of the two readouts. The paper's own panels show that.
Bcl6 and Stat3 were validated only as a dark-zone to light-zone ratio. There is no double-heterozygous multiome for those two, so a restored ratio does not show that centrocyte recycling returned. Even the Ctcf double-heterozygous multiome only partially reverses the pre-memory exit.
In silico knockout sets expression and motif activity to the minimum value in the dataset. It is not an allelic dose and not a time course. Jacobian edges and finite-difference edges disagree, and the paper plots both. Literature germinal-center edges are static, so a mismatch with a dynamic edge is not automatically a model error.
The velocity loss pulls the model toward scVelo. The authors say not to set that weight at 10,000 if scVelo already draws the wrong trajectories in the system at hand. Smooth latent time and accurate perturbations want different prior weights. Every number here is from mouse germinal centers, with no cross-tissue or cross-species test.