Interpretable models for transcriptional dynamics during cell fate transitions

Single-cell RNA sequencing experiments capture mixtures of cells at different physiological states, and enable the investigation of entire developmental trajectories. Current standard methods for fitting these trajectories typically rely on constructing cell-cell similarity graphs or high-dimensional functions, which have limited physical interpretability and do not provide directional information.

Seeking to construct firmer grounding for trajectory inference, we adapt concepts from stochastic biophysics and dynamical systems theory. To formalize the problem, we define a discrete model of RNA transcription and couple it to a sampling process. This approach yields a latent variable model, with RNA counts drawn from an occupation measure parametrized by kinetic rates. The latent variable model can be fit using the EM algorithm, providing biophysically meaningful parameters that characterize RNA production and processing during the developmental process. The framework generalizes to multimodal data, and reduces to current methods under particular simplifying assumptions. Compared to heuristic approaches, our model encodes a collection of concrete physical mechanisms, which facilitates statistical testing and provides guidance for further hypothesis-driven research.