EnsembleRegNet: Interpretable deep learning for transcriptional network inference from single-cell RNA-seq

Abstract

Gene regulatory networks (GRNs) govern gene expression and cellular identity, but accurately inferring their structure from high-dimensional single-cell RNA sequencing (scRNA-seq) data remains a major challenge. Here, we present EnsembleRegNet, a deep learning framework that infers transcription factor (TF)–target gene relationships by integrating an ensemble encoder-decoder and multilayer perceptron (MLP) architecture. EnsembleRegNet utilizes Hodges-Lehmann estimator (HLE)-based binarization, case-deletion analysis, motif enrichment using RcisTarget, and regulon activity scoring with AUCell to enhance both robustness and biological interpretability. Extensive evaluations across simulated and real scRNA-seq datasets demonstrate that EnsembleRegNet outperforms existing GRN inference methods, including SCENIC and SIGNET, in both clustering performance and regulatory accuracy. By uncovering cell-type-specific regulatory modules and enhancing interpretability, EnsembleRegNet offers a scalable and biologically grounded framework for exploring transcriptional regulation. Its demonstrated performance establishes a new benchmark for GRN inference and highlights its promise for applications in disease modeling, biomarker discovery, and cellular reprogramming.