Benchmarking Machine Learning Architectures for Gene Regulatory Network Inference

Details

Supervisors

Massart, Estelle

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enGene regulatory networks (GRNs) capture the regulatory interactions that govern gene expression, providing insight into how cellular identity is established and how dysregulation of these interactions can drive disease. This makes GRNs a valuable tool in drug discovery, where identifying the regulators responsible for aberrant gene expression can help prioritise therapeutic targets. Inferring these networks is difficult, as no complete experimentally validated reference of true regulatory interactions exists, making evaluation of GRN inference methods challenging. GRETA, a recently developed benchmarking framework, addresses this by comparing inferred networks through similarity and stability, and evaluating them against three complementary families of metrics capturing causal, predictive, and literature-based evidence. This thesis investigates whether deep learning can improve GRN inference by integrating two neural network-based methods, LINGER and DeepMAPS, into GRETA, and benchmarking them against existing inference methods on peripheral blood mononuclear cell data. The results suggest that deep learning does not systematically outperform existing methods. LINGER remains competitive across most evaluation tasks and its transcription factor activity module shows substantial agreement with SCENIC+, the current state-of-the-art method. DeepMAPS underperforms across most evaluation tasks, which may partly reflect the difficulties encountered in reproducing its original implementation.