Presentation Information
[C10-04]Inferring key regulators of cell fate via dynamical systems learning
*Masato Ishikawa1, Atsushi Mochizuki1 (1. Kyoto University (Japan))
Keywords:
dynamical system,gene regulatory network,single-cell RNA-seq,NeuralODE
Cell properties and functions are determined by gene expression dynamics, which emerge from gene-gene interactions. If a regulatory function (or differential equation model), which generates the expression dynamics, is obtained, various insights into the regulation of gene expression could be gained through analysis and numerical simulation of the regulatory function. Specifically, it will lead to discovering key genes that can control cell fate and identifying molecular mechanisms that regulate expression dynamics.
In this study, we developed NODEN, a computational method for inferring the regulatory function that generates expression dynamics based on gene regulatory networks and single-cell RNA-seq data. NODEN learns the regulatory function based on a neural network (NeuralODE) using the prior information on the regulatory relationships between genes obtained from the gene regulatory network. Accurate inference of regulatory functions requires accurate gene regulatory networks. Therefore, NODEN integrates multiple regulatory networks with different evidence, such as networks estimated by gene perturbations and networks estimated based on the binding of transcription factors to regulatory regions on the genome, in a data-driven manner.
Benchmarking with simulated data showed that NODEN improved the accuracy of regulatory networks by integrating multiple inaccurate regulatory networks, which resulted in more precise regulatory functions. Numerical simulations using the obtained regulatory functions identified combinations of key genes that control cell fate. Furthermore, NODEN was applied to real single-cell RNA-seq data of mouse pancreatic cells. Based on the inferred regulatory functions, NODEN predicted key transcription factors that control the differentiation of pancreatic endocrine cells (alpha, beta, and delta cells). These key factors were supported by multiple databases and literature.
In this study, we developed NODEN, a computational method for inferring the regulatory function that generates expression dynamics based on gene regulatory networks and single-cell RNA-seq data. NODEN learns the regulatory function based on a neural network (NeuralODE) using the prior information on the regulatory relationships between genes obtained from the gene regulatory network. Accurate inference of regulatory functions requires accurate gene regulatory networks. Therefore, NODEN integrates multiple regulatory networks with different evidence, such as networks estimated by gene perturbations and networks estimated based on the binding of transcription factors to regulatory regions on the genome, in a data-driven manner.
Benchmarking with simulated data showed that NODEN improved the accuracy of regulatory networks by integrating multiple inaccurate regulatory networks, which resulted in more precise regulatory functions. Numerical simulations using the obtained regulatory functions identified combinations of key genes that control cell fate. Furthermore, NODEN was applied to real single-cell RNA-seq data of mouse pancreatic cells. Based on the inferred regulatory functions, NODEN predicted key transcription factors that control the differentiation of pancreatic endocrine cells (alpha, beta, and delta cells). These key factors were supported by multiple databases and literature.