CellPace

CellPace is a generative model for single-cell developmental dynamics developed by researchers at McGill University and posted to bioRxiv in February 2026. Single-cell omics technologies capture only static snapshots of what are fundamentally continuous developmental processes, and because cells are destroyed during measurement, recovering coherent temporal trajectories is difficult—especially when developmental stages are sampled irregularly or some timepoints are missing entirely. CellPace is designed to reconstruct and extrapolate these dynamics from such incomplete time-series data.

The model's defining choice is to treat time as continuous rather than as a set of discrete categories. Many prior generative approaches encode each sampled timepoint as a separate class label, which hinders interpolation across temporal gaps and prevents extrapolation to future, unobserved stages. CellPace instead conditions a transformer-based diffusion backbone on continuous, gap-aware temporal encodings, allowing it to generate cell states at arbitrary points along a developmental axis. The authors describe the framework as a "temporal diffusion-forcing" approach, combining diffusion-based generation with a forcing scheme over time.

Evaluated across diverse mouse developmental lineages, CellPace reports state-of-the-art performance on three complementary tasks—simulation (generating realistic cell populations), interpolation (filling in missing intermediate timepoints), and forecasting (predicting future developmental stages). It joins a growing family of generative models for cellular dynamics such as scNODE and scDiffEq, but emphasizes irregular sampling and forward extrapolation as first-class capabilities.

#Key Features

• Continuous, gap-aware temporal conditioning: Time is encoded continuously rather than as discrete categories, enabling interpolation across missing timepoints and extrapolation to unobserved future stages.
• Three-in-one task coverage: A single framework handles simulation, interpolation, and forecasting of single-cell dynamics, rather than specializing in only one.
• Fine-grained biological fidelity: Generated cells preserve more than global population statistics—they retain dynamic gene regulatory programs and fine-grained cell-state structure.
• Multi-modal extension: The framework extends to joint RNA–chromatin dynamics, modeling coupled modalities even when temporal ordering is inferred from pseudotime rather than measured directly.

#Technical Details

CellPace uses a transformer-based temporal diffusion backbone conditioned on continuous, gap-aware temporal encodings, framed as a diffusion-forcing scheme over the developmental time axis. The model is trained and evaluated on diverse mouse developmental lineages, where it reports state-of-the-art results on simulation, interpolation, and forecasting relative to existing generative baselines. Beyond matching global population-level statistics, the authors show that generated cells preserve fine-grained biological structure, retaining dynamic gene regulatory programs across the trajectory. The framework also generalizes to multi-modal settings, jointly modeling RNA and chromatin accessibility dynamics; in cases where no explicit experimental time labels are available, temporal ordering can be supplied via pseudotime inference. As a recent preprint, specific parameter counts and full hyperparameter details are not summarized here, and the authors do not report publicly released weights at the time of posting.

#Applications

CellPace is aimed at developmental biologists and single-cell researchers who need to reconstruct continuous cellular trajectories from sparse or irregularly sampled time-course experiments. Its interpolation capability can recover plausible intermediate cell states between measured timepoints, while its forecasting capability can project lineages forward to developmental stages that were never sampled—useful for hypothesis generation about where a population is heading. The multi-modal extension supports integrated analysis of transcriptional and chromatin dynamics, relevant to studies of gene regulation during differentiation. Because the model preserves gene regulatory programs in its generated cells, it can also serve as an in silico simulator for designing or prioritizing future experiments.

#Impact

By modeling developmental time as a continuous variable, CellPace addresses a concrete and common obstacle in single-cell time-series analysis: irregular sampling and missing stages that defeat category-based generative models. Its unification of simulation, interpolation, and forecasting in one framework, together with multi-modal support, positions it as a flexible tool for reconstructing cellular dynamics. As a February 2026 preprint without yet-reported released weights, its real-world adoption and independent validation remain to be established, and broader testing beyond mouse developmental lineages will help clarify how well its continuous-time approach generalizes across systems.