Single-cell experiments capture cells frozen at a moment. The interesting biology is the motion between moments, and predicting where a perturbed cell will land has been the hard part. A team at Memorial Sloan Kettering Cancer Center, Weill Cornell Medicine and collaborating institutes in Essen and Barcelona built a model that learns that motion rather than assuming it.
Published in Nature Biotechnology, DynaVelo is a generative neural ODE model. From single-cell multiome data it learns how gene expression and transcription factor motif activity move together. Because it learns an actual vector field inside the model, researchers can perturb it in silico instead of only drawing trajectories through it.
Predicting the immune response
The team tested it on germinal center B cells, the structures where antibodies mature. It reproduced the documented split in wild-type mouse germinal center B cells, then mapped how those branches shift once one copy of the epigenetic regulators Arid1a or Ctcf disappears.
The interesting test came next. Simulated perturbation of the Arid1a-heterozygous cells pointed to Ctcf, Stat3 and Bcl6 as possible rescue targets. Experiments in double-heterozygous mice confirmed rescue of the Arid1a dark-zone phenotype for all three.
Networks that break where expected
The in silico perturbations also yielded state-specific gene regulatory networks. These recovered known transcription factor to gene edges in wild-type germinal centers and predicted which edges break in the mutants, then validated against multiome profiling of the double-heterozygous animals.
This is early work in mouse B cells, not a therapeutic program, and germinal center biology is a narrow slice of human immunology. What the paper does establish is that a learned dynamical model can nominate a short rescue list that survives in vivo testing, which is a far more demanding bar than reconstructing a trajectory plot.
