Ordinary differential equations to construct invertible generative models of cell type and tissue-specific regulatory networks

Abstract

1AbstractSingle-cell RNA-Seq (scRNA-seq) transcriptomics can elucidate gene regulatory networks (GRNs) of complex phenotypes, but raw sequencing observations only provide ”snap-shots” of data and are inherently noisy. scRNA-seq trajectory inference has been utilized to solve for the missing observations, but disentangling complex dynamics of gene-gene interactions at different time points from aggregated data is a non-trivial task and computationally expensive. Here we describe our Non-Stiff Dynamic Invertible Model of CO-Regulatory Networks (NS-DIMCORN) to define the genetic nexus underpinning specific cellular functions using invertible warping of flexible multivariate Gaussian distributions by neural Ordinary differential equations. Our results yield a generative model with unbiased density estimation from RNA-seq read-count data only. This resulted in scalable time-flexible sampling of each gene’s expression level thence allowing ab initio assembly of gene regulatory networks in specific cells. We demonstrate our proposed methodology is superior to the state-of-the-art algorithms in accurately recovering genome-wide functional interactions, whether from synthetic or empirical data. We optimized our algorithm for GPU-based implementation thereby further enhancing the utility of our proposed methodology in comparison to the ten benchmarked methods.