The paper introduces CellOT, a machine learning framework designed to predict how individual cells respond to biological stimuli, such as drugs or genetic changes. Because measuring a cell's molecular state typically destroys it, researchers cannot observe the same cell before and after a treatment; CellOT overcomes this "pairing problem" by using neural optimal transport to map unperturbed cell populations to perturbed ones. By employing input convex neural networks, the model learns the most likely transition trajectories for heterogeneous cell subpopulations based on a principle of minimal effort. The authors demonstrate that CellOT identifies nuanced, cell-state-specific drug responses and generalizes effectively to previously unseen patients and different species. Compared to existing autoencoder-based methods, this approach more accurately captures the variability and higher-order moments of complex cellular distributions. Ultimately, the research offers a scalable tool for understanding personalized medicine and the mechanistic drivers of cellular evasion in cancer therapies.

References:

• Bunne, C., Stark, S.G., Gut, G. et al. Learning single-cell perturbation responses using neural optimal transport. Nat Methods 20, 1759–1768 (2023). doi.org