Abstract
AbstractUnderstanding gene regulatory networks (GRNs) is crucial for unraveling cellular mechanisms and enhancing therapeutic interventions. Original methods for GRN inference from bulk expression data often struggled with the high dimensionality and inherent noise in the data. The prevalence of dropout events and background noise inherent in single-cell RNA sequencing has only increased the challenges. Here we introduce RegDiffusion, a novel neural network structure inspired by Denoising Diffusion Probabilistic Models but focusing on the regulatory effects among feature variables. Unlike other GRN methods for single-cell RNA-sequencing data, RegDiffusion introduces Gaussian noise to the input data following a diffusion schedule. It is subsequently trained to predict the added noise using a neural network with a parameterized adjacency matrix. We show that using this process, GRNs can be effectively learned with a surprisingly simple model architecture. In our benchmark experiments, RegDiffusion demonstrates superior performance compared to several baseline methods on multiple datasets. This work not only introduces a fresh perspective on GRN inference but also highlights the promising capacity of diffusion-based models in the area of single-cell analysis. The RegDiffusion software is available athttps://github.com/TuftsBCB/RegDiffusion.
Publisher
Cold Spring Harbor Laboratory