BEAR-GRN: Systematic assessment of single-cell multi-omics-based gene regulatory network inference methods

Abstract

Identifying regulatory gene interactions is inherently challenging as it depends on both the transcriptional and epigenetic landscapes. Multiple methods are proposed to infer gene regulatory networks (GRNs) using single-cell multi-omics data, yet the absence of a standardized benchmarking framework remains an obstacle to unbiased performance evaluation. Here we present BEAR-GRN, a systematic resource providing standardized datasets, curated ground-truth networks, and a unified computational pipeline for benchmarking GRN inference methods. Using BEAR-GRN, we assess state-of-the-art multi-omics GRN inference methods across paired single-cell RNA-seq and ATAC-seq datasets spanning human and mouse cell types. We evaluate accuracy against four complementary ground-truth definitions, alongside computational efficiency, output stability, and modality contribution. Although ground-truth choice profoundly influences performance, LINGER and DIRECT-NET show the highest overall accuracy and stability. Modality perturbation experiments reveal that current methods are predominantly RNA-driven, with chromatin accessibility contributing limited independent signal. BEAR-GRN enables researchers to evaluate existing and new GRN methods against established baselines. BEAR-GRN provides a standardized framework for benchmarking gene regulatory network inference methods using single-cell multi-omics data, revealing that ground-truth choice profoundly impacts performance and that current methods remain predominantly RNA-driven.