Cell fate gets an inverse design loop: LabCompass lifts erythroid progenitors to 22% in four lab rounds
Synopsis
Culture conditions combine continuous doses, so testing them exhaustively is impractical, and earlier work could only rank predefined candidate designs or search a discrete catalogue. LabCompass couples a flow-matching generative model with a forward surrogate that predicts cell phenotypes, using the surrogate's gradient at inference time to steer sampling toward recipes expected to produce a target cell state. In primary human cord-blood CD34+ cells, four experimental loops produced a 78-million-cell spectral flow cytometry dataset, recovered a known megakaryocyte-inducing recipe the model had never seen, and raised erythroid progenitors to 22%, roughly 1.6-fold over the control condition.
Interpretation
The search over culture conditions moves from a fixed candidate catalogue to directly generated recipes, with the target cell state steering generation from the first step. Earlier cellular phenotype optimization largely relied on retrieval or discrete search, ranking predefined candidate designs by their ability to induce a target state, so it could not generate new designs and handled continuous doses poorly. On a synthetic grid-optimization problem, LabCompass was compared with uniform random search across all tested dimensionalities and resolutions; at resolution 5 the relative performance rose monotonically from 1.73 in 2 dimensions to 97.91 in 30.
Generation starts from a flow that gradually reshapes random noise into realistic recipes, and at each step the forward surrogate evaluates what cell composition the current candidate would induce. Flow matching is a generative method that learns how to continuously transform noise into a data distribution; here it acts as a prior over recipes, keeping optimization inside experimentally plausible regions. In primary cord-blood CD34+ cells, 116 designs spanned 10 cytokines or small molecules, two oxygen levels and culture days, and a 20-color spectral flow cytometry panel profiled 50 million cells.
Experimental measurements are fed back so that each loop updates both the recipe prior and the forward surrogate. The first loop held out six designs, including design 210, a literature-known condition yielding about 70% megakaryocyte progenitors, while no training design exceeded 16%. Without having observed design 210, the model generated recipes close to its UM171 at 70 nM and butyzamide at 100 nM, and predictive uncertainty correlated negatively with Euclidean distance to design 210 (Spearman ρ = −0.72).
Perspective
The demonstrated setting is in vitro culture of primary human cord-blood CD34+ cells, with readout defined by a 20-color spectral flow cytometry immunophenotype; the authors also apply the same framework to single-cell transcriptomic data from HCMV infection and suggest it could extend to other stem-cell differentiation systems, organoid cultures and drug screening.
The model can only reach target states supported by the available design space: monocytes and early lymphoid populations were enriched only after M-CSF and a lymphoid cocktail were added in the third loop, while pHSC and Pre-ProB enrichment stayed limited under the initial parameters. The authors also note that performance depends on the quality and diversity of training data, that the deep learning model offers limited mechanistic interpretability, and that retraining the forward surrogate each loop adds computational burden.