Depicting pseudotime-lagged causality across single-cell trajectories for accurate gene-regulatory inference

Abstract

Identifying the causal interactions in gene-regulatory networks requires an accurate understanding of the time-lagged relationships between transcription factors and their target genes. Here we describe DELAY, a convolutional neural network for the inference of gene-regulatory relationships across pseudotime-ordered single-cell trajectories. We show that combining supervised deep learning with joint-probability matrices of pseudotime-lagged trajectories allows the network to overcome important limitations of ordinary Granger causality-based methods, such as the inability to infer cyclic relationships such as feedback loops. Our network outperforms several common methods for inferring gene regulation and predicts novel regulatory networks from scRNA-seq and scATAC-seq datasets given partial ground-truth labels. To validate this approach, we used DELAY to identify important genes and modules in the regulatory network of auditory hair cells, as well as likely DNA-binding partners for two hair cell cofactors (Hist1h1c and Ccnd1) and a novel binding sequence for the hair cell-specific transcription factor Fiz1. We provide an open-source implementation of DELAY at https://github.com/calebclayreagor/DELAY.

Dynamics of chromatin accessibility during human first-trimester neurodevelopment

“…4d). We next applied DELAY 25 , a different convolutional neural network method that exploits the temporal shift between the expression of transcription factors and their targets in single-cell lineages in combination with ChIP-seq derived TF binding site information to estimate gene regulatory networks (GRNs). This revealed a network of 148 TFs co-regulating each other during Purkinje cell differentiation (Fig.…”

“…Supervised inference and stochastic simulation of Purkinje GRN. We used DELAY 25 (https://github.com/calebclayreagor/DELAY) to infer the Purkinje GRN from gene-accessibility dynamics in pseudotime then performed stochastic simulations to verify the putative network's gene-expression dynamics. First, we re-trained DELAY on a large scATAC-seq dataset of plasma B-cell differentiation 61 with ChIP-seq ground-truth data 62 to prepare the neural network to infer the Purkinje GRN from tens of thousands of single nuclei.…”

Gene regulatory network inference with popInfer reveals dynamic regulation of hematopoietic stem cell quiescence upon diet restriction and aging

“…These are built on the premise that the higher-resolution o ered by decomposing bulk samples into single cells can improve GRN inference, and that the single-cell noise does not overwhelm the signal. These include methods rooted in statistical learning [23], dynamical systems theory [24], treebased approaches [25], information theory [26,27,28], and time series analysis [29]. More recently, methods also consider dynamic changes to network topology itself [30].…”