U-Pert

Understanding how a perturbation — a drug, a cytokine, or a genetic edit — reshapes a population of cells is central to both basic biology and therapeutic design. Most computational perturbation models focus on how gene expression shifts within cells, but perturbations also change how many cells of each type there are: they drive proliferation, death, and differentiation that reshape the composition of the population itself. U-Pert (Unbalanced Perturbation Dynamics For Cell Fate Design) is a single-cell model that learns both effects at once, jointly capturing transcriptomic state transitions and cell-number ("mass") dynamics from single-cell snapshots.

U-Pert was introduced in 2026 by Qiangwei Peng, Yuchuan Wang, Jianzhe Li, Xinyu Wang, Yao Xiao, and Peijie Zhou at the Center for Machine Learning Research, Peking University — a group whose work centers on unbalanced-dynamics and optimal-transport modeling of single-cell trajectories. The "unbalanced" framing is the key distinction: whereas balanced transport methods assume cell number is conserved between conditions, U-Pert allows mass to be created and destroyed, so the model can attribute observed changes in a population to a combination of shifting cell states and changing cell abundances.

Beyond fitting observed data, U-Pert is built for two forward-looking tasks. It predicts the cellular response to perturbations and contexts not seen during training, and it performs inverse design — searching for interventions that would steer a population toward a desired outcome, the "cell fate design" of its title. This places it alongside perturbation-response predictors such as CPA and GEARS while adding an explicit account of population growth and depletion.

#Key Features

• Joint state-and-mass modeling: U-Pert models transcriptomic state transitions and cell-number dynamics together, so predicted responses reflect both how cells change and how the abundance of each cell type shifts under perturbation.
• Unbalanced dynamics from snapshots: By relaxing the mass-conservation assumption of balanced optimal transport, the model learns growth and death from unpaired single-cell snapshots rather than requiring lineage tracing or paired before/after measurements.
• Forward prediction of unseen perturbations: Once trained, U-Pert predicts responses to perturbations and biological contexts absent from the training data rather than being refit on each new query.
• Inverse design of interventions: The framework searches over perturbations to identify those most likely to produce a target cell-state distribution, supporting hypothesis generation for cell-fate engineering.
• Multi-setting evaluation: The model is assessed on controlled simulations, genetic perturbation benchmarks, the sciPlex3 drug-response dataset, and PBMC cytokine perturbations, spanning genetic, chemical, and signaling interventions.

#Technical Details

U-Pert is a generative dynamics model grounded in the unbalanced optimal transport framework, which extends classical optimal transport with a growth term so that probability mass — here, cell number — need not be preserved between observed conditions. From single-cell snapshots that are not paired across time, the model infers how cells move through transcriptomic space and how sub-populations expand or contract, yielding a joint description of molecular and abundance change. This design directly targets a limitation of expression-only perturbation predictors, which can misattribute a change in observed cell proportions to a change in cell state. The authors evaluate the approach across four settings — controlled simulations with known ground truth, genetic perturbation benchmarks, the sciPlex3 chemical screen, and PBMC cytokine perturbations — reporting that it recovers both molecular and abundance changes.

#Applications

U-Pert is aimed at researchers studying how cell populations respond to interventions, including drug and cytokine screening, functional genomics, and developmental or immune cell-fate studies. Its forward-prediction mode can prioritize which perturbations to test experimentally by predicting responses to unseen conditions, while its inverse-design mode proposes interventions expected to drive cells toward a desired state — useful in cell engineering and regenerative-medicine contexts where the goal is to reprogram or expand a target population. Because it explicitly models cell-number changes, it is especially suited to settings where a perturbation alters population composition, not just per-cell expression.

#Impact

By coupling transcriptomic state transitions with cell-number dynamics, U-Pert broadens single-cell perturbation modeling from "how do cells change?" to "how does the whole population change?", addressing a blind spot in expression-only methods. It extends the Peking University group's line of unbalanced optimal-transport research toward the virtual-cell goal of predicting and designing cellular responses in silico. U-Pert is a preprint released under a CC BY-NC-ND license and is awaiting peer review, and no code or pretrained weights have yet been released, so independent benchmarking against established perturbation predictors remains to be reported.