Proteins that share specific biological functions are expected to be interconnected within a PPI network. Therefore, the primary purpose of PPI networks is to unravel the functions of unidentified proteins by using the annotations of known genes. In addition, the network structure information can facilitate addressing biological questions, including the identification of hub proteins, novel pathways, and evolutionary analysis of proteins of interest (von Mering et al., 2002). Noteworthy, the link between two proteins has various implications, such as altering the kinetic properties of enzymes, affecting the substrate binding affinity of effectors, and modifying the regulatory effects of TFs on their downstream target genes (Berggard et al., 2007). Given that transcription regulation depends on both TFs and their associated cofactors, PPI networks offer supplementary insights into transcription regulation (Serebreni and Stark, 2021). This additional layer of information is distinct from the link between TF and regulatory DNA elements.

Currently, only a limited number of experimental-based proteomic networks have been established in plants (Consortium, 2011). Using a yeast-two-hybrid (Y2H) mapping workflow, a comprehensive map of the phytohormone signaling network was constructed, revealing the multifaceted functions of phytohormone proteins in Arabidopsis (Altmann et al., 2020). Protein mass spectrometry was also used to identify protein complexes in 13 plant species (McWhite et al., 2020). Recently, Han and colleagues conducted a Y2H screening of 7,623 baits against 21,964 prey proteins in maize, resulting in the identification of 56,243 high-confidence PPIs by vigorous filtering (Han et al., 2023). Moreover, there are computational algorithms have been developed for PPI prediction (Zhang et al., 2016a). A support vector machine (SVM) model has been trained to generate a Protein-Protein Interaction Database for Maize (PPIM), covering ~ 2,700,000 interactions among ~ 14,000 proteins (Zhu et al., 2016). Yang and coworkers have established a comprehensive database (PlaPPISite) for 13 plant interactomes by collecting experimentally validated PPIs and computational predictions (Yang et al., 2020).

Despite these significant advances, well-explored plant PPI maps remain limited. Current data shows that the available PPI datasets cover ~ 12,000 genes in Arabidopsis, ~ 40 genes in Soybean, and ~ 300 genes in Rice based on the Biogrid database (accessed on 20 April 2024) (Oughtred et al., 2021). Therefore, it is important to establish a larger quantity of high-quality PPI networks through coordinated efforts by the research communities in the future.

With the rapid accumulation of transcriptomics data, such as gene expression microarrays and RNA-seq data, GCNs are frequently employed to elucidate the connection between genes and to cluster a large number of genes that exhibit similar expression patterns (Stuart et al., 2003). GCNs represent indirect connections without considering directionality. They are typically generated by a weighted network construction approach followed by hierarchical clustering to identify smaller co-expression modules (Langfelder and Horvath, 2008). While we can use prior knowledge of TF-coding genes to assign the directionality from TFs to their target genes, the directionality between two TFs always remains unknown in the GCN. Despite the lack of causal regulatory links, mounting evidence suggests that GCNs are efficient in predicting the specific biological functions of unknown genes by the “guilt-by-association” principle (Wolfe et al., 2005) and in identifying hub genes that exhibit high connectivity with other genes and may have important regulatory roles (Lin et al., 2019).

To pinpoint regulatory or functional genes involved in specific biological processes, functional modules associated with various pathways or traits are partitioned from large GCNs and annotated by Gene Ontology (GO) terms enrichment analysis. For example, a GCN was constructed and divided into 25 modules in wheat. These modules were annotated and connected to the spatiotemporal progression during wheat endosperm development (Pfeifer et al., 2014); Co-expression modules were also identified for secondary biosynthetic pathways in tea plants (Tai et al., 2018).

As GCNs inherently lack information regarding regulatory relationships among co-expressed genes, it is necessary to combine co-expression analysis with additional complementary data sources, such as cis-regulatory data. Integration approaches can enhance the reliability of GCNs for capturing true biological relevance from network connections. By integration of co-expression data, cis-regulatory elements, and conserved DNA motifs, Vandepoele and coworkers were able to accurately link many unknown genes to specific biological functions, such as the E2F pathway in Arabidopsis (Vandepoele et al., 2009).

In plants and other organisms, TFs regulate structural genes and TF-coding genes with high context specificity (Badia-i-Mompel et al., 2023). GRN analysis serves as a robust tool for delineating the regulatory relationships between a single or a set of TFs with distinct functions and their downstream target genes in specific cellular and environmental conditions. It has also shown its value in identifying key regulator TFs, regulatory connections between genes and pathways, and in formulating testable functional and regulatory hypotheses (Van den Broeck et al., 2020).

GRNs can be classified into two groups based on their objectives: context-dependent GRNs and comprehensive untargeted GRNs (Depuydt et al., 2023). The majority of GRN studies have been designed to elucidate the network wiring that underlies specific developmental processes or responses to particular environmental conditions. For example, a Bayesian-based network analysis was used to identify multiple genes associated with the SHOOT MERISTEMLESS TF gene and to predict their roles in shoot apical meristem development (Scofield et al., 2018). Borrill and colleagues integrated time-series data in wheat and identified several hub genes, including the well-known senescence regulator NAM-A1, which regulates the expression of senescence-related genes within the network (Borrill et al., 2019). Zander and coworkers generated a GRN model to predict the cross-talk in the jasmonic acid (JA) signaling pathway and to discover novel components involved in the JA regulatory mechanism (Zander et al., 2020). Furthermore, known and novel candidate TFs were identified associated with water-deficit responses and xylem development plasticity using integrative network analysis in rice (Reynoso et al., 2022).

While context-dependent GRN studies often provide high-resolution information on the specific biological process under investigation, untargeted GRN approaches, despite having lower resolution, are able to capture a broader range of biological processes under various conditions. Untargeted GRNs are typically generated using extensive datasets without focusing on only one specific biological question. Instead, they have been used to establish a database resource or test novel algorithms. For example, Zhou and colleagues collected extensive transcriptome datasets to create coexpression-based GRNs in maize (Zhou et al., 2020). Recently, several resource articles have been published, such as MaizeNetome (Feng et al., 2023), Wheat-RegNet (Tang et al., 2023), and wGRN (Chen et al., 2023b). To introduce more different context-specificities, it is common to incorporate a lot of complementary datasets from various tissues, treatments, and developmental stages. Moreover, the integration of additional omics layers, such as trait-association results, can provide further evidence for the hypotheses drawn from the transcriptome and identify more accurate candidates for the following experimental validation (Kim et al., 2023). Nevertheless, although these GRNs are very large, containing millions of edges, they are not saturated yet.

Single-cell omics technologies, particularly single-cell RNA sequencing (scRNA-seq), provide comprehensive insights into the transcription landscape of diverse plant tissues, surpassing conventional bulk sequencing methods (Rhee et al., 2019). As gene regulation principally takes place in individual cells, inferring regulatory networks based on single-cell data is more effective than using bulk data. It predicts interactions based on expression within the same cells rather than averages (Chen et al., 2019). Moreover, the increased resolution of single-cell omics data allows to capture the cell type- or state-specific GRNs (Aibar et al., 2017).

Current single-cell assays are limited in their ability to detect all transcripts in every cell, often capturing fewer than 5,000 genes per cell. Therefore, specialized tools have been developed to handle this data sparsity (Hao et al., 2024). Common network inference methods designed for scRNA-seq data include SCODE (Matsumoto et al., 2017), PID (Chan et al., 2017), Inferelator (Jackson et al., 2020), and SCENIC/SCENIC+ (Aibar et al., 2017; Van de Sande et al., 2020; Bravo Gonzalez-Blas et al., 2023). These methods vary in their underlying models for linking regulator TFs to target genes. SCENIC first identifies regulatory relationships based on co-expressed genes using GENIE3 (Huynh-Thu et al., 2010) or GRNBoost2 (Moerman et al., 2019), and then refines the connections by considering TF binding motifs on promoter regions. The defined “regulons” consist of co-expressed genes enriched for the CREs to which the regulatory TF binds. Finally, the workflow identifies cells where these regulons are active (Aibar et al., 2017). However, the lack of validated and formatted TF-DNA binding data for most plant species hinders the application of these methods with plant scRNA-seq data.

Single-cell technologies now allow for the quantification of many other modalities, such as scATAC-seq (Buenrostro et al., 2013). GRN methods have been developed to combine the data from multiple modalities (Jiang et al., 2022a; Alanis-Lobato et al., 2024), or alternatively, networks can be constructed separately with each modality and then integrated together (Jansen et al., 2019). Nevertheless, unlike bulk sequencing technologies, which capture a higher number of transcripts, the sparsity inherent in single-cell data may result in biased estimations of gene expression correlations (van Dijk et al., 2018). We expect these challenges to be addressed through enhanced sequencing depths and more sophisticated bioinformatics methodologies to effectively manage data with limited counts (Sekula et al., 2020; Song et al., 2023).

Currently, compared to PPI network and GCN, GRN has emerged as a favored tool for predicting essential regulators and gene expression alterations in response to environmental stimuli and intrinsic signals (Gupta et al., 2022). In some articles, broadly defined GRNs can be formed by the connections between regulatory elements that regulate the transcriptional and translational processes. Such elements, including TFs, splicing factors, and microRNAs, could be incorporated into the modeling of GRNs (Lai et al., 2016; Carthew, 2021) ( Figure 1 ). In the following sections, we will use the term “GRN” to refer to the network that abstracts the directed relationships between TFs and their target genes in the context of transcriptional regulation and emphasize studies related to GRNs.

The edges in a physical network represent interactions between a TF and the specific CREs of the target genes it regulates (Schmitz et al., 2022). It is important to note that the physical interaction edges do not imply a functional alteration in gene expression. Instead, they represent a regulation potential that contributes to the complex transcription process.

TFs bind to the genomic TF binding sites (TFBSs) through their DNA binding domains (DBDs). DBDs typically recognize short DNA motifs. Both experimental and computational approaches have been used to identify TFBS and DNA motifs recognized by specific DBDs. Large amount of TFBS and motif datasets have been collected and deposited in public databases, such as TRANSFAC (Matys et al., 2003), CIS-BP (Weirauch et al., 2014), JASPAR (Rauluseviciute et al., 2024), UniPROBE (Hume et al., 2015), PlantPAN (Chow et al., 2024), and ChIP-Hub (Fu et al., 2022).

A functional regulatory network is characterized by TF-target edges that are supported by changes in the expression patterns of the target genes. These connections, whether direct or indirect, can reflect the functional impact of the regulator’s actions on their targets. From bulk RNA-seq data to single-cell transcriptomics, advanced inference methods have been developed, demonstrating enhanced accuracy and computational efficiency (Marbach et al., 2012a; Pratapa et al., 2020). Several approaches, which utilize time-series data or pseudo-temporal single-cell transcriptomics, were designed to gain more precise insights into the regulatory interactions between genes (Huynh-Thu and Geurts, 2018; Aubin-Frankowski and Vert, 2020).

Several statistical approaches are used for the transcriptome-based gene network analysis. The underlying principles of these methods include correlation, supervised learning, probabilistic models, dynamical-systems modeling, and deep learning (Li et al., 2015; Kim et al., 2023).

In line with the concept of combining different methods in predicting functional GRN, combining physical and functional interactions datasets is also essential to construct comprehensive and high-confidence regulatory networks. The integration process can be achieved by simply using the ChIP-seq data for a TF along with the matched RNA-seq data in the mutant or by employing more advanced algorithms to merge the information from various multi-omics datasets.

One should bear in mind that the connections in a GRN are hypothetical and require vigorous evaluation of their accuracy. Several common practices have been established to evaluate the biological relevance of the inferred connections ( Li et al., 2015 ).

The most common evaluation method involves comparing the inferred GRN with a “gold standard” network, which is often derived from experimentally verified results, such as loss and gain of function experiments. These wet-lab approaches generate confident regulatory connections by observing the impact of a regulator’s expression changes on its target gene ( Kim et al., 2023 ). When performing the comparison, one may calculate the average accuracy according to the detection ratio of the verified edges, which is probably flawed due to the sparsity of GRNs. In other words, an algorithm that always predicts the absence of edges could incorrectly achieve high accuracy. Thus, a better approach is to assess the proportion of correctly identified positives relative to all positives (sensitivity or recall) and the proportion of correctly identified positives out of all identified positives (precision or positive predictive value) (Huynh-Thu and Sanguinetti, 2019).

When experimentally validated “gold standard” or any well-accepted high-confidence networks are unavailable, alternative approaches for evaluating gene networks may include cross-validation tests and functional coherent module assessment. Cross-validation tests the accuracy of the reconstructed network by predicting gene functions based on the known functions of network neighbors. Additionally, high-quality networks are expected to exhibit coherent modules of interacting and co-regulated genes. The functional coherence of these modules can be evaluated through enrichment tests of gene function and probabilistic models to predict gene expression within the module (Li et al., 2015).

In addition to connecting TFs and their target genes, GRNs can provide further insights into gene functions and associated biological processes through downstream analysis.

Early-stage studies often relied solely on bulk gene expression data, typically obtained from specific plant tissues or organs, to construct functional GCNs.

For example, Pfeifer and colleagues analyzed gene expression in developing wheat grains and constructed a co-expression network comprising 25 modules. These modules displayed unique spatiotemporal characteristics that can be distinguished based on grain cell types or developmental stages (Pfeifer et al., 2014). To provide insights into the coordination of individual homoeologs underlying various traits in wheat, the coexpression networks were constructed from nonstress tissue-specific and stress-related RNA-seq samples. These networks highlight the extensive coordination of homoeologs throughout development and in response to various stresses and offer a platform to identify candidate genes for agronomic traits (Ramírez-González et al., 2018). Huang and coworkers evaluated various parameters for data normalization and different inference methods for constructing a large GCN in maize using RNA-Seq data. The analysis revealed that increasing sample size positively impacts network performance, emphasizing the importance of sample size for the construction of accurate GCNs (Huang et al., 2017). To extend the knowledge of salt response in soybean, Hu and colleagues clustered differentially expressed genes between a salt-tolerant and a salt-hypersensitive cultivar. They constructed undirected networks representing their co-expression patterns based on Pearson’s correlation coefficients. The network analysis unveiled several candidate pathways critical in salt responses, including phytohormone signaling, oxidoreduction, phenylpropanoid biosynthesis, and others (Hu et al., 2022).

While GCNs cannot directly define the regulatory relationships between TFs and their downstream targets, the connectivity results are widely used to refine the findings from genome-wide association studies (GWAS) and quantitative trait locus (QTL) analyses. This integration enables the effective prediction of novel candidate genes. For example, a weighted GCN analysis was used to identify connected genes associated with Fusarium head blight (FHB) resistance and pinpointed candidate hub genes within the interval of three previously reported FHB resistance QTL in wheat (Sari et al., 2019). Yao and coworkers combined GWAS and co-expression network analyses to uncover candidate genes involved in the accumulation of oleic acid content in rapeseed (Yao et al., 2020). GCN analysis and genome-wide association studies (GWAS) were combined to elucidate the regulatory pathways and identify candidate genes responsible for pre-harvest sprouting and seed dormancy traits in maize (Ma et al., 2023b).

Establishing direct physical interactions between TF and DNA has been a major research focus. In addition to TFBSs obtained from experimental techniques such as ChIP-seq, many TFBSs have been predicted by computational algorithms. For example, Yu et al. collected transcriptomics data from developing maize leaves and used co-expression data along with enrichment analysis to predict overrepresented motifs in the promoter sequences and the potential TFBSs of key TFs (Yu et al., 2015).

A few databases and newly developed inference methods have significantly expanded the available information on TF-DNA binding interactions. PlantPAN, which has collected a comprehensive set of public ChIP-seq datasets, is a valuable resource for plant TF-TFBS interactions. It offers the most complete plant PWMs for analyzing TFBSs and effective tools for predicting TFBSs in conserved regions of a given promoter. The latest version, PlantPAN 4.0, provides a non-redundant set of 3,428 matrices for 18,305 TFs of 115 plant species (Chow et al., 2024). Another valuable resource is ChIP-Hub, a comprehensive and standardized platform for exploring the regulome of plants. It collects over 10,000 datasets from 41 plant species and processes them based on ENCODE standards. As an application example, an extensive survey was performed to examine the co-associations among various regulators, enabling the construction of a hierarchical regulatory network spanning a broad developmental context (Fu et al., 2022).

Meanwhile, wet-lab approaches persist in being actively employed to extend experimental TF-DNA binding in plants. In wheat, Zhang and colleagues have successfully obtained high-quality DNA binding profiles for 53 environmentally responsive TFs using DAP-seq. Interestingly, the study found that 85% of the in vitro identified TFBSs were located within transposable elements and associated with regulatory sequences specific to the wheat lineage (Zhang et al., 2021). In a subsequent study by the same group, genomic binding profiles were generated for a larger set of TFs, enabling the assembly of a wheat GRN encompassing connections among 189 TFs and 3,714,431 regulatory elements (Zhang et al., 2022b). These results provide valuable insights into the transcriptional regulatory mechanisms in wheat. Several remarkable advances were also made in maize. Ricci and coworkers performed DAP-seq on 32 TFs, indicating that the distal accessible chromatin regions were enriched for TFBS (Ricci et al., 2019). Additionally, interaction maps were generated for 14 maize TFs from the ARF family, revealing both specific and redundant binding events of ARF TFs (Galli et al., 2018). Furthermore, 104 maize TFBS datasets were yield by ChIP-seq with transient expressed proteins to construct the leaf regulatory network (Tu et al., 2020).

GRNs derived from gene expression profiles are not limited by the availability of TF-DNA binding data and are widely used in various biological contexts. A GRN was inferred by modeling 78 maize seed transcriptome to identify key genes involved in seed development. The network analysis unraveled highly interwoven communities and identified key genes and regulatory modules associated with nutrient transport and imprinting patterns, which are crucial for maize seed development (Xiong et al., 2017). Utilizing the GENIE3 software package with a number of RNA-Seq data, Huang and colleagues constructed four tissue-specific GRNs in maize. They further predicted key TFs for each specific tissue (Huang et al., 2018).

Zhou and colleagues present a standardized pipeline using machine learning algorithms along with transcriptomic data to predict GRNs (Zhou et al., 2020). They analyzed a large collection of transcriptome datasets, resulting in 45 GRNs. The networks exhibited significant enrichment for biologically relevant interactions, with each GRN capturing diverse biological processes. This uniform pipeline can be applied to other species with available expression data (Zhou et al., 2020). To comprehensively elucidate the chloroplast biogenesis process, Loudya and colleagues present a biologically informed GRN. The network prediction suggests that the regulators of chloroplast genes are differentially involved across various leaf developmental stages in wheat (Loudya et al., 2021).

Similar to GCN, GRN can also be combined with QTL and GWAS results to predict candidate genes for specific traits. For example, Zhao and colleagues designed an integrative analysis combining eQTL, GWAS, and GRN to characterize the genetic basis of cotton yield. Several high-ranking causal genes identified from the GRN were validated for their functional impacts on cotton seed development (Zhao et al., 2023).

Genomics and functional genomics studies on rice have been at the forefront among crop plants. Rice also serves as a leading model in integration network studies. The RiceNet (v2) web resource, launched in 2014, provides an integrative network for rice. This network combined co-functional links based on genomic context similarity, connections inferred from co-expression patterns, and protein-protein interactions. Its utility in prioritizing candidate genes involved in rice biotic stress responses has been demonstrated (Lee et al., 2015a). Another significant pioneering study created a comprehensive developmental atlas of maize with multi-omics data. Integrative GRNs were constructed based on mRNA, protein, and phosphor-protein data, resulting in improved predictive power. This work enhanced our understanding of the complex regulatory mechanisms in maize (Walley et al., 2016).

The integration of multi-omics data has become increasingly prevalent in studies using network-based approaches. Han and colleagues have successfully constructed a large-scale PPI network in maize. An integrated map was constructed incorporating data from four different layers: three-dimensional genomics, transcriptomics, proteomics, and protein-protein interactions. Leveraging this multi-omics network and machine learning-based prediction approaches, novel candidate key genes involved in various regulatory pathways, such as flowering time, have been predicted and genetically validated (Han et al., 2023). Gomez-Cano and coworkers analyzed ~4.6M interactions, including co-expression networks, TF-DNA interaction experiments, and expression quantitative trait loci (eQTL) to construct GRNs and pinpointed key regulators associated with hormone, metabolic, and developmental processes (Gomez-Cano et al., 2024). Additionally, several studies integrate a large amount of omics data, including both physical interactions and functional regulation relationships, in wheat (Chen et al., 2023; Tang et al., 2023). Similar integrated analysis has also been conducted in cotton (Zhao et al., 2024).

The integration of network and genetic mapping data, such as GWAS, further enhances the predictive power for identifying significant genes. Lin and colleagues thoroughly examined the transcriptome and epigenome profiles of the developing spike in an elite wheat cultivar. Through the integration of regulatory networks with GWAS, key genes affecting the spike architecture were pinpointed (Lin et al., 2024).

Regulatory networks relying on bulk data have several limitations. These models typically only capture generalized connection patterns, which obscure distinct regulatory interactions unique to certain cell types. Furthermore, bulk data often fails to differentiate the cellular states, which can significantly impact gene regulation. In contrast, approaches such as microdissection and single-cell technologies enable the discovery of regulatory networks at greater spatial resolution.

Zhan and coworkers used laser-capture microdissection with RNA-Seq to profile gene expression in each dissected cell compartment of the maize kernel (Zhan et al., 2015). They constructed an unbiased GCN and detected sub-network modules containing genes predominantly expressed in a single compartment or ubiquitously expressed across multiple compartments. These results offer a high-resolution gene expression atlas of maize kernel and contribute to uncovering regulatory interactions associated with the differentiation of major endosperm cell types (Zhan et al., 2015).

With the advent of sing-cell omics, Marand and colleagues generated a cis-regulatory atlas using single-cell ATAC-seq in maize. They profiled over 72,000 nuclei across six maize organs and identified TFs coordinated with chromatin interactions by analyzing patterns of co-accessible CREs. This comprehensive cis-regulatory atlas at single-cell resolution is a valuable resource to study the gene regulation in maize (Marand et al., 2021). The researchers from Vandepoele’s group have developed computational methods named MINI-EX and MINI-AC to explore cell type-specific regulatory interactions. MINI-EX utilizes expression-based GRNs derived from single-cell RNA-seq data and TF binding motifs to predict cell type-specific regulons. On the other hand, MINI-AC combines accessible chromatin (AC) data from either bulk or single-cell experiments with TF binding motifs to construct GRNs (Ferrari et al., 2022; Manosalva Perez et al., 2024). The application of MINI-EX has successfully identified regulons (groups of genes co-regulated by a shared TF) across major cell types in Arabidopsis, rice, and maize. Moreover, this method effectively prioritized established key regulons based on their network characteristics, such as connectivity and centrality, and unveiled several previously unidentified transcriptional regulators (Ferrari et al., 2022). Similarly, MINI-AC has also demonstrated superior performance compared to other techniques in accurately identifying TFBS. Maize has a complex genome and abundant distal AC regions. MINI-AC successfully inferred leaf GRNs containing experimentally confirmed interactions between TFs and target genes from both proximal and distal regions in maize. It is also a robust tool for pinpointing both known and novel candidate regulators (Manosalva Perez et al., 2024).

Recently, Yuan and coworkers focused on the differentiation stage of maize endosperm. They performed single-cell RNA-seq combined with TFBS profiling using ampDAP-seq to construct a high-confidence GRN and identified key regulators in five distinct cell types (Yuan et al., 2024). Fu and colleagues utilized the endosperm spatial transcriptome data during the grain-filling stage. They successfully predicted and identified the function of the candidate sucrose transporter genes (SUTs) in endosperm transfer cells facilitated by GCN analysis (Fu et al., 2023).

Unlike traditional forward genetics, a new research paradigm is emerging for gene function studies, wherein candidate genes are determined through hints from networks. Y1H is an easy approach used to identify the direct binding between TFs and the promoters of their targets. These direct regulatory relationships have great value in guiding the selection of key regulators for functional characterization.

For instance, Gaudinier and colleagues used enhanced Y1H assays to screen for Arabidopsis TFs binding to the promoters of genes associated with nitrogen metabolism and signaling, resulting in a network comprising 1,660 interactions. This network unveiled a hierarchical regulation of these TFs. Mutants of 17 prioritized key TFs exhibited significant alterations in at least one root architecture trait. The identification of regulatory TFs in the nitrogen-regulatory framework holds promise for enhancing agricultural productivity (Gaudinier et al., 2018). Similarly, Shi and coworkers uncovered TFs that regulate genes related to mycorrhizal symbiosis using Y1H. They screened more than 1500 rice TFs for binding to 51 selected promoters, and constructed a highly interconnected network. Interestingly, many of the TF in this network are involved in the conserved P-sensing pathway. With functional analyses of selected genes, this study elucidates the extensive regulation of mycorrhizal symbiosis by both endogenous and exogenous signals (Shi et al., 2021).

Ji and colleagues constructed a co-expression network to identify regulatory factors during the grain-filling stage of maize endosperm, and identified hundreds of candidate TFs using 32 storage reserve-related genes as guides. In addition to known regulators of storage proteins and starch, the study uncovered novel TFs, such as GRAS11, involved in endosperm development. They further characterized the function of GRAS11 through detailed functional analysis (Ji et al., 2022). High-temporal-resolution RNA sequencing was conducted on the basal and upper regions of maize kernels. Weighted gene co-expression network analyses were performed, identifying numerous hub regulators that are worthy of subsequent functional characterization (He et al., 2024).

Collectively, these studies have provided significant insights into transcriptional regulation programs and rich data resources. GCN analyses have identified modules and candidate genes associated with various traits. Experimental determination of TFBSs, aided by computational predictions, has enabled the construction of regulatory networks, revealing novel regulators. Integration of multi-omics data has improved the predictive power of GRNs. High-resolution spatial techniques have uncovered cell-type-specific regulatory interactions, providing a more nuanced understanding of gene regulation. Overall, the advancements in regulatory network studies of crop species have substantially enhanced our understanding of the complex transcriptional programs governing plant growth, development, and responses to biotic and abiotic stresses.

The validation of regulatory networks is crucial to ensure that these networks accurately reflect the biological processes of interest. GRN evaluation commonly requires a thorough comparison of predicted interactions with the “gold standard” derived from wet-lab experiments, such as independently generated TF-DNA binding data and perturbation tests on the regulator TFs, as discussed in Section 4.1. However, the experimental “gold standard” is often unavailable or inadequate. Noteworthy, even if TF binding to a target is confirmed by in vivo ChIP assay, it does not necessarily imply that this TF can activate or repress the target gene. An alternative approach to test whether a TF regulates a particular gene is to perturb the TF expression and check how this perturbation affects the expression of target genes. While this approach shows promise due to its inherent causality, perturbation experiments are time consuming and costly. Noteworthy, they are likely not to work well as hindered by the widely existed compensatory mechanisms in crop plants. Due to different validation methods have their own limitations, utilizing diverse assessment strategies to evaluate a given GRN may be a smart way.

Regulation of gene expression is a dynamic process. High-temporal resolution studies have revealed fluctuations in gene expression levels in maize kernels within a small time window (Yi et al., 2019). From the perspective of network inference, there may be lack of expression correlation between target genes and their regulatory TFs because of the temporal lag between TF binding and the accumulation of mRNA transcripts. Thus, improving the spatiotemporal resolution of gene expression profiles and TF-DNA binding data is imperative for network construction.

Although new technologies have been developed to more efficiently acquire multi-omics data from a single plant cell (Xu et al., 2022b), the primary challenge arises from technical difficulties in the preparation of high-quality protoplasts in plants. Recently, to address challenges in protoplasting experiments, several optimized enzymatic cell wall digestion protocols have been developed for various species (Ye et al., 2022; Wang et al., 2022b; Chen et al., 2023a). Wang and colleagues introduced a new method that involves two consecutive digestion processes with different enzymatic buffers, significantly enhancing the efficiency and viability of protoplast preparation across diverse plant tissues (Wang et al., 2022b). However, the conversion of cells into protoplasts is still not feasible for many types of plant tissues. Alternatively, recent single-nucleus techniques offer broader applicability across different tissues. Nevertheless, it’s important to note that nuclear RNA and cytoplasmic RNA should not be considered equivalent.

There is still a large gap between the knowledge derived from GRNs and their manifestation in agronomic traits. Firstly, GRNs involve a complex interplay among thousands of genes and TFs that underlie various biological processes. Deciphering how perturbations in these regulatory networks impact gene expression remains a challenge. Secondly, the relationship between gene expression levels and traits is nonlinear and polygenic. Therefore, predicting observable traits from changes in gene expression, especially considering the influence of environmental factors, is difficult.

At the current stage, it is feasible to modulate specific metabolic pathways based on network information. The activity of one or a few TFs can regulate multiple steps of metabolic pathways. Thus, manipulating the expression of TFs probably has a greater impact on metabolism pathways than modifying cis-regulatory elements of enzyme-coding genes. For example, flavonoids are considered valuable compounds in plant metabolic engineering. Increasing flavonoid levels can be achieved by manipulating their transcription regulatory elements, resulting in the development of plants with high anthocyanin content (Jiang et al., 2023). For the enhancement of oil production, WRINKLED1 is a conserved transcription factor involved in the regulation of fatty acid biosynthesis in diverse angiosperms. Transgenic plants that overexpress the WRINKLED1 gene show promising outcomes in increasing the oil content of maize, soybean, and rice. As WRINKLED1 also modulates targets that affect plant growth and development. It is important to consider the shared regulatory network when utilizing it to engineer plant oil production (Yang et al., 2022).

However, achieving the modulation of complex traits, such as yield and quality, which are determined by multiple factors, remains challenging. The regulatory mechanisms that directly impact these processes have not been thoroughly characterized. And these complex traits are often influenced by environmental factors and are sensitive to the interplay between genotype and environment.

In response to current challenges, there are several aspects in future network-related research that need to be strengthened. Firstly, integrating more omics data can enhance the predictive power of networks by merging diverse complementary information. Secondly, improving spatiotemporal resolution relies on the development of more sensitive, convenient, and cost-effective technologies. Lastly, the application of deep learning models, which can better integrate massive amounts of data and extract reliable and useful information from them, provides an opportunity to construct more accurate GRNs.