Prioritizing Maize Metabolic Gene Regulators through Multi-Omic Network Integration

Gene regulatory networks (GRNs) link transcription factors (TFs) to the biological processes they control. Assembling them remains difficult because the relevant data types (gene expression, protein-DNA interactions, and genetic variation) are large, heterogeneous, and rarely combined. Here, we developed and benchmarked a framework that integrates these data into TF-function predictions in maize. We assembled four complementary TF-target gene network layers, based on expression, protein-DNA interaction, trans-expression quantitative trait loci (eQTL), and cis-eQTL-supported interaction, from 46 Random Forest (RF)-inferred regulatory networks, 283 protein-DNA interaction assays, and eQTLs derived from 16 million SNPs across 304 inbred lines. Together these layers comprised ~4.6 million interactions. We then compared three strategies for integrating them, benchmarking each against published TF knockout data. A network-based approach, which represents every gene as a low-dimensional vector (embedding) learned from the combined network, outperformed the two overlap-based strategies, annotating over eight times more TFs (~3,000), agreeing most closely with gene knockout responses where predictions existed, and remaining robust when individual layers lacked data. The predictions recovered TF functions and predicted new regulators of hormone, developmental, and metabolic processes, which we prioritized per process and mapped to specific conditions. Using similarity on the low-dimensional vector representation (embedding), we further identified candidate functionally redundant or diverged TF paralogs. Because it relies only on data types now common across species, the framework provides a generalizable template for prioritizing regulatory genes in maize and other plants.