scNODE : Generative Model for Temporal Single Cell Transcriptomic Data Prediction

Abstract

Temporal measurements using single-cell RNA sequencing (scRNA-seq) technology enable the study of gene expression programs of individual cells for normal developmental processes or diseased physiological states. However, due to the labor, expense, and technical challenges associated with these experiments, researchers are only able to profile gene expression at discrete and sparsely spaced time points. This results in information loss between consecutive discrete time points and impedes the downstream developmental analysis. To address this problem, we proposescNODE, an end-to-end model that can simulate and predict realisticin silicosingle-cell samples at any time point to enable temporal downstream analyses. Our method integrates a variational autoencoder (VAE) with neural ordinary differential equations (ODEs) to predict gene expression in a continuous and non-linear latent space. Importantly,scNODEadds a regularization term to integrate overall dynamics of cell developments to the latent space, such that the learned latent representation is informative and interpretable. Our evaluations on six real-world scRNA-seq datasets show thatscNODEachieves higher predictive performance than state-of-the-art methods. We further demonstrate thatscNODElearns an interpretable latent space, and helps recapitulate cell-type developmental trajectories.AvailabilityThe data and code are publicly available athttps://github.com/rsinghlab/scNODE.

Multi-condition and multi-modal temporal profile inference during mouse embryonic development

“…"-would be immensely valuable in understanding the temporal dynamics and regulation of the behaviors of individual cells across biological conditions. [23] [42,37,6,11,51,50] [13] [14] Table 1: Methods for integrating and extrapolating single-cell time series data. OT = optimal transport; ODE = ordinary differential equations; AE = autoencoder.…”

“…For example, optimal transport-based methods, such as Waddington-OT [36], TrajectoryNet [42] and TIGON [37], operate on pairs of measurements in a time series, with the goal of aligning cells from two neighboring time points under the assumption that a single cell's profile changes minimally between time points. Alternatively, neural ordinary/stochastic differential equation (ODE/SDE) based methods, such as PRESCIENT [49], RNAForecaster [6], MIOFlow [11], FBSDE [51] and scNODE [50], assume that each cell develops autonomously, and the cell's future state is determined based on the cell's current expression profile. Autoencoder models either assume that time is an additive variable in the embedding space [23] or are coupled with optimal transport or ODE based methods to optimize non-linear cell projection and cross-time alignment [42,11,50].…”

“…Alternatively, neural ordinary/stochastic differential equation (ODE/SDE) based methods, such as PRESCIENT [49], RNAForecaster [6], MIOFlow [11], FBSDE [51] and scNODE [50], assume that each cell develops autonomously, and the cell's future state is determined based on the cell's current expression profile. Autoencoder models either assume that time is an additive variable in the embedding space [23] or are coupled with optimal transport or ODE based methods to optimize non-linear cell projection and cross-time alignment [42,11,50]. Graph-based methods, such as GraphFP [13], perform dynamic inference on top of cell graphs.…”