Benchmarking of Graph Autoencoder models for Gene Regulatory Network inference with prior knowledge

Abstract

This thesis explores the benchmarking of graph auto-encoders for inferring gene regulatory networks (GRNs) using prior knowledge of known GRNs. Gene regulatory networks are crucial for understanding gene expression and their regulation in various biological processes. However, obtaining experimentally validated GRNs can be expensive and time-consuming. In this thesis, we propose using graph auto-encoders, a type of neural network, to learn the underlying structure of GRNs from gene expression data. We evaluate different types of graph auto-encoders, including the standard Graph Auto-Encoder (GAE), Variational Graph Auto-Encoder (VGAE), Adversarially Regularized Graph Auto-Encoder (ARGA), and Adversarially Regularized Variational Graph Auto-Encoder (ARGVA) to infer GRNs from prior knowledge. Implementation of a trainable decoder shows better results in comparison to the standard. We compare the performance of each autoencoder with respect to the performance and stability of inferred GRNs. This study demonstrates the potential of different architectures of graph auto-encoders for inferring gene regulatory networks using prior knowledge to compare performance and stability in gene network inference.