Gene regulatory network inference using deep learning-based feature scores

Abstract

Reconstructing gene regulatory networks (GRNs) from expression data remains a central challenge in computational biology. While tree-based ensemble methods such as GENIE3 and GRNBoost2 have long been the gold standard, deep learning offers the potential to capture more complex regulatory patterns, provided its predictions can be made interpretable. This thesis investigates whether feature-importance mechanisms embedded in deep learning architectures can be exploited for GRN inference. Two methods embedding feature scores in their architectures, LassoNet and TabNet, are integrated into a GENIE3-like per-gene regression framework, and a stability-selection procedure is used for LassoNet to aggregate feature-importance scores across resampled regularization paths into more robust interaction ranking. These methods are benchmarked against GENIE3 and GRNBoost2 but also against other deep learning-based methods such as DeepSEM or scGeneRAI on the DREAM and BEELINE datasets using AUPRC and Early Precision Ratio. Results show tree-based ensembles retain a clear advantage on synthetic and smaller curated networks, while somehow trailing DeepSEM in experimental and bigger networks. No single method dominates uniformly across biological contexts and network scales.