Climate change is already affecting ecosystems around the world and forcing us to explore ways to adapt to meet societal needs. This is particularly true in crop science where researchers are working to identify and predict genes and their resulting phenotypes under different environmental conditions in order to secure food production under a new climate regime (Thudi et al., 2021; Tian et al., 2021). Understanding gene/phenotype/environment relationships requires a large data set which can be difficult to collect, so most researchers focus on a small number of heavily studied species. The speed with which climate change is progressing necessitates a massive scaling up of the number of species with understood G/P/E dynamics. The research and the knowledge gained in this area will also help human exploration in space, where plants will play an important role (Barker et al., 2023). Previous study has demonstrated that knowledge about one species can be applied to another using ontologically supported knowledgebases that exploit homologous structures and orthologous genes (Naithani et al., 2020). These types of knowledge structures that can apply knowledge about one species to another have the potential to enable the massive scaling up that is needed.

An important part of predicting phenotype is understanding the complex gene regulatory networks present in plants. This study will focus on the promoter region, the 5′ cis-regulatory regions of the homologs. This region is a portion of the DNA strand that is “upstream” from the 5′ end of the gene's coding start site and provides selective binding sites for trans-acting factors such as transcription factors, repressors, and activators that regulate the expression of the gene (Liu et al., 1999). These regions are just one element of the gene expression process. Studying the expression of trans-acting factors is important for understanding the spatiotemporal dynamics of molecular interactions that help adapt or overcome stress. Resources such as the Gene Ontology (The GO Consortium, 2021), Planteome (Cooper et al., 2018), Plant Reactome (Naithani et al., 2020), and KnetMiner (Hassani-Pak et al., 2021) contain much of what we know about gene function, gene regulatory networks, and phenotypes in the form of Gene X regulates Gene Y and Gene Y impacts phenotype Z, but the contextual effect of environmental conditions under which these interactions happen is almost always not included in the annotations. Not all plants and their genes are characterized in detail, but if it is included, the environmental context is usually detailed only in the metadata. Investigations that use protein domain identification and gene homology-based methods to infer the functional role a gene carries out in a given species may be overlooking the spatial and temporal dynamics of mRNA expression that determines whether a gene product (protein) will be present at the desired time and place to serve a molecular function. The interactive nature of genes, environments, and phenotypes requires a data structure that can represent qualitative relationships (e.g., “has phenotype” or “regulates”) and integrate heterogeneous data types in a single, queryable framework. One of these data structures is a knowledge graph (KG) (Sheth et al., 2019).

A graph is made up of objects (nodes) and the relationships (edges) between those objects and, in this context, represents what we know about how biological and environmental entities (objects) interact. Rather than store data in a table or database, a knowledge graph stores the synthesized knowledge we gain from the data, e.g., Gene X has phenotype Y. As more knowledge is added to the graph, more complex queries, network analyses, and inferences can be made. Important examples include the use of knowledge graphs in rare disease diagnosis in humans (Zemojtel et al., 2014), drug repurposing (Reese et al., 2021), improving cancer treatment (Gogleva et al., 2022), and meta-analyses (Tiddi et al., 2020). KGs used for translational science rarely contain environmental exposures even though we know environmental conditions are an important part of gene expression dynamics. The exact way to model exposures in a KG is still under development (Chan et al., 2023). A KG containing information about plant genomics and phenomics under different environmental conditions can be used to generate hypotheses in silico for targeting, thereby reducing the number of in vivo experiments that need to be conducted, saving time and resources.

This study examines gene expression patterns in response to drought conditions in four plant species, such as Arabidopsis thaliana, Zea mays, Sorghum bicolor, and Populus trichocarpa. The central motivation of this study is to assess the feasibility of using homologs to make predictions about gene expression in multiple species.

The graph query returned 62 pairs of homologous genes from Sorghum bicolor, Zea mays, Arabidopsis thaliana, and Populus trichocarpa (Supplementary material 6), but only 16 pairs between A. thaliana and P. trichocarpa had documented similar (8) and differential (8) expressions in drought conditions (Table 3). All of the genes with similarly expressed pairs had decreased expression. Expression data for the 16 homologous pairs of A. thaliana and P. trichocarpa came from two studies in GXA (de Simone et al., 2017; Filichkin et al., 2018).

Based on the predicted transcription factor-binding sites in the promoter regions, the Populus genes in the differentially expressed homolog pairs cluster separately from the other Populus and Arabidopsis genes (Figure 3). This difference is driven by a group of 11 transcription factor-binding sites that are absent in the promoter regions of the subset of divergent Populus genes (RAV, MIKC, NAM, G2-like, CPP, ARR-B, tify, TALE, NF-YC, ERF, and NF-YA). The separation of these genes cannot be explained by the taxon or the study providing the data (which overlaps the taxon).

There were seven Populus genes that clustered away from the others. All but one (POPTR_014G056800v3 involved in strigolactone biosynthesis) were hypothetical proteins (According to Gramene). GO annotations for these genes clustered around transporter activity, catabolic activity, response to stress, binding, and catalytic activity. The 11 transcription factors absent in the binding sites of the Populus genes include proteins involved in plant stress response in Arabidopsis (According to UniProt).

A comparison of the promoter regions between homolog pairs showed that homologs that were expressed similarly had more similar promoter regions than pairs that were expressed differentially (Figure 4).

Separate comparisons of the promoter regions from gene pairs with contrasting expression profiles also show that gene pairs with similar expression had more similar promoter regions (Figure 5A) and gene pairs with contrasting expression had less similar promoter regions (Figure 5B).

This study shows that one can use in silico experiments to predict gene expression in drought conditions using homologous gene families in some species pairs but not all. This study supports previous findings that in some cases, promoter regions evolve separately from the coding region of the genes they regulate (Tirosh et al., 2008). Thus, we can translate knowledge about gene expression in one species to another, but we need to include these dynamics in the data infrastructures we use to make this translation, in this case, KGs. Many data structures link a gene to a phenotype, trait, or disease without specific expression information. The current representation of differential gene expression links exposure to a chemical or a drug to the increased or decreased expression of a specific gene in the context of toxicology and drug development (Fecho et al., 2022; Unni et al., 2022). Gene regulatory networks are represented as mini-networks of genes that influence other genes (The GO Consortium, 2021), but many of these networks are still unknown in plants. In the short term, in silico KG experiments involving gene expression can be improved by including empirically validated gene expression patterns of homologs.

Gene regulatory networks in plants have been developed using a combination of experimental and computational approaches (Kulkarni and Vandepoele, 2020). Methods combining high-throughput DNA sequencing (ChIP-seq) and expression data have successfully revealed the detailed regulatory networks controlling flowering (Chen et al., 2018) but are difficult to scale. Methods such as ATAC-seq and DAP-seq are more scalable but only reveal a partial picture of the regulatory network (O'Malley et al., 2016; Maher et al., 2018). KGs can be used to infer regulatory networks at scale, but the quality is highly dependent on the data used to build the KG. The advantage of applying a KG is the ability to integrate incredibly heterogeneous data in a single graph, thus modeling regulatory networks in their larger biological context. An example of this application is the relatively new field of “network medicine” that uses KGs to examine the progression of disease (Silverman et al., 2020). The main disadvantage of KGs in this application is that large amounts of computable data and domain-specific knowledge models are needed to create a graph of this type. Many disciplines do not have these resources available. While KGs can infer gene regulatory networks, these networks should always be confirmed using established experimental and computational approaches.

These opposing gene expression patterns are not a concern for researchers who are only interested in finding a list of genes that are potentially important in a specific context. It is not until one needs to generate hypotheses about the impact of the environment on the biological function that more complex graph representations become needed. If we are to incorporate the effect of the environment, we need to know more than that Gene X has phenotype Y. We need to know if the environmental effect increases or decreases the expression of the gene and the biological consequences of that change in expression. In some cases, we may only know that an environment is linked to a specific phenotype without knowing the underlying mechanism. This information can still add useful knowledge to the graph. In some cases, the graph itself can be used to generate hypotheses about the interplay between genes, biological processes, molecular functions, cellular components, and an observed phenotype.

Despite having the graph available to quickly explore the data and locate genes of interest, the workflow for comparing the promoter regions required substantial manual intervention. In this instance, we only had 16 gene pairs to explore, but scaling up these types of analyses will require the ability to traverse data annotated with gene identifiers and gene coordinates. Future studies should include extending the graph model to include these data types.

The semantic representation of the effect of environmental exposure on gene expression is more straightforward for the effects of a chemical or a substance, such as phenol or rubber cement. Data can be collected in the laboratory using model organisms, and the results added to the graph for analysis and translational research. Everyday environmental exposures are rarely this simple and frequently involve exposure to many types of substances in different contexts, such as climate or socioeconomic status. Future studies may need to develop ontologies and semantic representations for these more complex exposures.

Our observations support our hypothesis and justify the extension of our KG to include TF-binding site annotations and the actual TF genes, which are either known empirically or are supported by co-expression network analysis. In future, an investigation of conservation vs. non-conservation of cis- and trans-regulatory regions of genes may improve the understanding of interspecies and intraspecies responses to stress and adaptation.

The merged KG data are hosted on the CyVerse DataCommons (https://datacommons.cyverse.org/browse/iplant/home/shared/genophenoenvo). The KG data are available for direct download or for remote visualization via CyVerse WebDav service (https://data.cyverse.org/dav-anon/iplant/commons/community_released/genophenoenvo/kg/) using visualization software such as Neo4J. The python code used to create the graphs are publicly hosted on GitHub (https://github.com/genophenoenvo/knowledge-graph). The final merged KG includes two tab-separated value (tsv) files which include the edges and nodes.

AT developed and framed research question(s), analyzed data, contributed to data analysis, developed software, contributed to writing and revising the paper, and project administration and management. HH contributed to data analysis. TS contributed to data analysis and contributed to writing and revising the paper. JR developed software, validated results or software, developed and framed research question(s), and contributed to writing and revising the paper. LC contributed to data analysis, project administration and management, and contributed to writing and revising the paper. PJ developed and framed research question(s), analyzed data, contributed to data analysis, and contributed to writing and revising the paper. JE contributed to data analysis, developed software, validated results or software, and contributed to writing and revising the paper. All authors contributed to the article and approved the submitted version.