Transcriptomic forecasting with neural ODEs

Abstract

Single-cell RNA-seq (scRNA-seq) technologies can uncover changes in the molecular states that underlie cellular phenotypes. However, to understand dynamic cellular processes, computational tools are needed to extract temporal information from the snapshots of cellular gene expression that scRNA-seq provides. To address this challenge, we have developed a neural ordinary differential equation based method, RNAForecaster, for predicting gene expression states in single cells for multiple future time steps. We demonstrate that RNAForecaster can accurately predict future expression states in simulated scRNA-seq data. We then show that using metabolic labeling scRNA-seq data from constitutively dividing cells, RNAForecaster accurately recapitulates many of the expected changes in gene expression during progression through the cell cycle over a three day period. Finally, we extend RNAForecaster to use unspliced and spliced counts from scRNA-seq to predict the impact of knock down experiments in pancreatic development. Thus, RNAForecaster enables estimation of future expression states in biological systems.

Biologically informed NeuralODEs for genome-wide regulatory dynamics

“…We tried to emulate how one might typically use OOTB NeuralODE models for the purpose of predicting gene expression dynamics [6]. Given a gene regulatory network of n genes, we assume that the gene expression of all genes g j (t) can have an effect on a specific g i (t): dg(t) dt = f reg g(t) − g(t), where g(t) = {g i (t)} n i=1 and f reg : R n → R n .…”

“…Given that many dynamical systems can be described using ordinary differential equations (ODEs), a logical approach to modeling GRNs is to estimate ODEs for gene expression using an appropriate statistical learning technique [3][4][5][6]. Although estimating gene regulatory ODEs ideally requires time-course data, obtaining such data in biological systems is difficult (if not impossible given the destructive nature of the associated assays).…”

“…These latter models can be broadly placed into two classes based on the inputs required to estimate the gradient f of the gene regulatory dynamics, where . The first class consists of methods like PRESCIENT [4] and RNAForecaster [6] that can learn f based only on time series gene-expression input without additional steps or consideration of other regulatory inputs [4, 6, 8]. In the process of learning transitions between consecutive time points, these “one-step” methods implicitly learn the local derivative (often referred to as “RNA velocity” [10]) , as an intermediary to estimating f .…”

“…In the process of learning transitions between consecutive time points, these “one-step” methods implicitly learn the local derivative (often referred to as “RNA velocity” [10]) , as an intermediary to estimating f . One significant issue with these approaches is scalability, and studying meaningfully large dynamical systems (ideally those describing the entire genome) has been too costly in terms of runtime and performance loss [4, 6, 9, 20]. This leads to potential issues with generalizability as regulatory processes operate genome-wide and even small perturbations can have wide-ranging regulatory effects.…”

“…While the flexibility of both classes of models is helpful in estimating arbitrary dynamics, they are”black-box” methods whose somewhat opaque nature not only makes them prone to overfitting, but it makes it difficult to extract interpretable mechanistic insights about regulatory control [1, 6]. These models are optimized solely to predict RNA velocity or gene expression levels and so the predictions are not explainable in the sense that they cannot be related back to a sparse causal GRN [16].…”