DynaVelo predicts three genetic interventions that restore a mouse immune-cell pattern

An AI model of changing immune-cell states has produced predictions that researchers could test in living animals. DynaVelo nominated three genetic interventions to counter an abnormal pattern in mouse germinal centres, where B cells refine the antibody response. Experiments supported all three, moving the work beyond reconstruction of familiar cell types to the harder task of suggesting which perturbation to try next.[1]

In a study published in Nature Biotechnology, the team trained a neural differential-equation model on measurements of gene expression and how accessible regulatory DNA is inside individual immune cells. Their test environment was the germinal center, the part of an immune response where B cells mature. The model learned the expected paths between proliferating dark-zone cells and differentiating light-zone cells. It also identified how those paths shifted in mouse cells carrying reduced function of genes such as Arid1a and Ctcf, which influence the packaging and regulation of DNA.[1]

Simulating a cell is not the same as predicting a label

Ordinary classification asks what kind of cell a measurement resembles. DynaVelo attempts to infer the direction and speed of change in an underlying biological state. It uses an artificial neural ordinary differential equation: a learned mathematical rule that describes how a compressed representation of gene expression and regulatory activity might evolve. That makes simulated changes possible. Researchers can decrease the modeled activity of a gene and examine whether the system moves toward a healthier reference pattern.[1]

The method does not observe a single cell continuously through time. The time direction is reconstructed from single-cell datasets and modeling assumptions, a source of uncertainty even when the inferred trajectories look biologically plausible. The technical achievement is to integrate RNA abundance with clues about transcription-factor activity from accessible DNA, rather than treat every gene as a disconnected marker. This helps represent the regulatory programs that steer a cell toward one fate rather than another.[1][2]

The team's mutant germinal-center cells did not behave like normal reference cells. Some characteristic circulation between dark and light zones was disrupted, changing the distribution of immune-cell states. DynaVelo predicted that manipulating three regulators—Ctcf, Stat3 and Bcl6—could counter part of the Arid1a-related abnormality. This was the essential experimental wager: would interventions suggested in a computer correspond to a measurable change in an animal?[1]

The answer was more substantial than a successful software benchmark. The researchers combined partial loss of Arid1a with partial loss of individual candidate genes in mouse models and examined germinal-center outcomes. For the newly generated combinations involving Stat3 and Bcl6, the abnormal dark-zone-to-light-zone ratio shifted back toward wild-type levels. The study also reports Ctcf-related rescue, with additional single-cell analyses supporting the predicted change in cell dynamics. In this constrained setting, the model anticipated genetic interactions that experiments could verify.[1]

Translating that result into medicine would require a different set of tests. Reducing a gene’s activity in an engineered mouse is not the same intervention as giving a drug, and a change that helps one cell population may harm another. DynaVelo’s immediate value is as a way to prioritise experimental targets and make their predicted effects explicit.[1]

That creates a useful feedback loop. Molecular measurements shape the model; the model nominates a perturbation; the experiment shows whether the predicted state change occurs. When a prediction fails, that failure can expose a missing relationship or an assumption that the model has treated too confidently.[1]

The three mouse results make that loop tangible. Researchers did not stop at drawing trajectories through molecular data. They selected genetic combinations, tested them and compared the resulting cells with the model’s expectations.[1]

Report an error or suggest a correction