scTour: a deep learning architecture for robust inference and accurate prediction of cellular dynamics

Abstract

Despite the continued efforts to computationally dissect developmental processes using single-cell genomics, a batch-unaffected tool that is able to both infer and predict the underlying dynamics is lacking. Here, I present scTour, a novel deep learning architecture to perform robust inference and accurate prediction of the cellular dynamics in diverse processes. For inference, scTour can efficiently and simultaneously estimate the developmental pseudotime, intronic read-independent vector field, and transcriptomic latent space under a single, integrated framework. For prediction, scTour can precisely reconstruct the underlying dynamics of unseen cellular states or an independent dataset agnostic to the model. Of note, both the inference and prediction are invariant to batch effects. scTour’s functionalities are successfully applied to a variety of biological processes from 16 datasets such as cell differentiation, reprogramming and zonation, providing a comprehensive infrastructure to investigate the cellular mechanisms underpinning development in an efficient manner.

UniTVelo: temporally unified RNA velocity reinforces single-cell trajectory inference

“…under the steady-state equilibrium to be both assigned in steadystate, instead of forcibly dividing into induction and repression stages as done by previous framework. Overall, this top-down framework enjoys computational convenience of both modelling the spliced RNAs with diverse distribution families, including deep neural networks as demonstrated in ScTour 25 and aggregating latent time across genes (see below), while preserving the same level of accuracy in a vanilla setting.…”

“…It aligns with a recently proposed framework MultiVelo to use chromatin accessibility to predict the transcription rate 29 . Also, this top-down design brings convenience to model spliced RNAs with a broader family of dynamic functions, including deep neural networks that have high complexity, and to predict the velocity even entirely without unspliced RNAs 25 .…”

UniTVelo: temporally unified RNA velocity reinforces single-cell trajectory inference

“…This benefits the RNA velocity quantification based on more reliable spliced RNAs and allows cells that are either above or under the steady-state equilibrium to be both assigned in steadystate, instead of forcibly dividing into induction and repression stages as done by previous framework. Overall, this top-down framework enjoys computational convenience of both modelling the spliced RNAs with diverse distribution families, including deep neural networks as demonstrated in ScTour [25] and aggregating latent time across genes (see below), while preserving the same level of accuracy in a vanilla setting.…”

“…It aligns with a recently proposed framework MultiVelo to use chromatin accessibility to predict the transcription rate [29]. Also, this top-down design brings convenience to model spliced RNAs with a broader family of dynamic functions, including deep neural networks that have high complexity, and to predict the velocity even entirely without unspliced RNAs [25].…”

Inferring single-cell transcriptomic dynamics with structured latent gene expression dynamics

“…These dynamics are considered "unstructured" in the sense that f is a general dense feed-forward neural network, and interactions between all components of z are allowed. A very similar model with these unstructured dynamics was also recently developed in scTour [22]. However, these unstructured models do not utilize biological knowledge of the the dynamics of splicing.…”