Interpretable and tractable probabilistic models for single-cell RNA sequencing

We outline a class of models of technical and biological variability for single-cell RNA sequencing (scRNA-seq). This technology can quantify the number of RNA molecules in millions of cells at a genome-wide scale. Due to the volume of the data, typical scRNA-seq analyses are ad hoc and descriptive, typically relying on normalization, background subtraction, and data filtering. We argue that the descriptive worldview is insufficient, as it oversimplifies the intrinsic variability in the transcriptional and experimental processes.

Our approach emphasizes a class of processes that are particularly suited to the analysis of scRNA-seq data. To facilitate biological discovery, these processes are interpretable: each parameter encodes the rate of a biophysical phenomenon. To enable use with large datasets, they are also tractable, translating the convolution of noise sources into the composition of generating functions. We present computationally facile solutions to generic biological systems, which couple stochastic transcriptional driving to downstream RNA processing. In addition, we outline probabilistic strategies for phenomena particular to scRNA-seq, such as non-ergodicity in developmental trajectories, background contamination, and variability in capturing individual molecules.