Trajectory inference from single-cell genomics data with a process time model

Single-cell transcriptomics experiments provide gene expression snapshots of heterogeneous cell populations across cell states. These snapshots have been used to infer trajectories and dynamic information even without intensive, time-series data by ordering cells according to gene expression similarity. However, while single-cell snapshots sometimes offer valuable insights into dynamic processes, current methods for ordering cells are limited by descriptive notions of “pseudotime” that lack intrinsic physical meaning. Instead of pseudotime, we propose inference of “process time” via a principled modeling approach to formulating trajectories and inferring latent variables corresponding to timing of cells subject to a biophysical process. Our implementation of this approach, called Chronocell, provides a biophysical formulation of trajectories built on cell state transitions. The Chronocell model is identifiable, making parameter inference meaningful. Furthermore, Chronocell can interpolate between trajectory inference, when cell states lie on a continuum, and clustering, when cells cluster into discrete states. By using a variety of datasets ranging from cluster-like to continuous, we show that Chronocell enables us to assess the suitability of datasets and reveals distinct cellular distributions along process time that are consistent with biological process times. We also compare our parameter estimates of degradation rates to those derived from metabolic labeling datasets, thereby showcasing the biophysical utility of Chronocell. Nevertheless, based on performance characterization on simulations, we find that process time inference can be challenging, highlighting the importance of dataset quality and careful model assessment.Author summary: Single-cell RNA sequencing can measure the amounts of RNA in individual cells, and although it is a snapshot experiment, cells that are differentiating can be captured in distinct states allowing for inference of “trajectories” or “velocity”. Currently, methods that attempt to do so rely heavily on heuristics, with no mechanistic meaning associated with the “pseudotime” they assign to cells. We show that it is possible to infer trajectories under a biophysical model within a principled framework. By developing a trajectory model based on cell state transitions, we demonstrate that it is possible to infer interpretable latent variables, i.e. process time, corresponding to the timing of cells subjected to a biophysical process, as well as transcriptional parameters with biophysical meaning. However, we find this to be a challenging task. By characterizing failure scenarios in simulations and with quantitative assessment on real datasets, we concluded such inference is not always possible, especially when there is insufficient dynamical information embedded in the data. In such cases, our trajectory model allows us to perform model selection to determine if captured cells are better modeled by clusters. Our findings emphasize the importance of thoughtful experimental design and meticulous model assessment for valid trajectory inference.