Plant metabolites generated via diverse biochemical pathways are major sources of nutrients [1], flavors [2], and medicine [3]. They also play important roles in interspecific interactions [4] and ecological adaptation [5]. In addition, knowledge of the enzyme genes responsible for producing and modifying plant metabolites is important for crop improvement [6] and novel therapeutic applications [7]. However, there are challenges in ascribing metabolites and enzyme genes to plant metabolic pathways. First, there is a large number of metabolites in each plant species, for example, the estimate for Arabidopsis is w5,000 [8]. Second, as opposed to general metabolism involving pathways conserved across plant species, many metabolites are produced in a taxa-specific (specialized) manner [9]. Thus, knowledge of specialized metabolic pathways in one species may not be transferable to another one. Third, enzyme genes tend to belong to large gene families with complex evolutionary histories [10]. Fourth, enzyme genes responsible for specialized metabolism are frequently homologs of general metabolism genes [11]. Finally, pathways intersect with each other to form a network and individual pathways can be challenging to delineate [12]. As a result, enzyme genes responsible for making metabolites, particularly specialized ones, are largely unknown, let alone their memberships in metabolic pathways. Despite these challenges, over the last few decades, we have accumulated substantial knowledge of metabolic pathway genes through biochemical and genetic approaches [13,14]. With the ever-increasing availability of biological data, candidate genes responsible for general and specialized metabolism can be predicted computationally, facilitating the discovery of new metabolic pathways. For example, enzyme genes can be assigned into pathways with additional knowledge of substrates or metabolic products and reactions that catalyze the substrate-product conversions [15]. Pathway memberships of genes can also be hypothesized based on coexpression profiles [16,17], co-localization in cellular compartments [18], or genes forming metabolic gene clusters (MGCs) [19]. More recently, statistical and mathematical modeling approaches have been used for resolving questions related to metabolic pathways, such as predicting the presence of a specific metabolic pathway [20], identifying metabolites [21], assigning metabolites or enzyme genes to pathways [17,22,23], distinguishing generalized and specialized metabolism genes [24], and modeling the metabolic flux [25]. Two recent reviews have also provided overviews of computational approaches for identifying specialized metabolic pathway memberships, including those based on sequence similarity, MGCs, gene co-expression, and genome-wide association study (GWAS) [26,27]. Here, we aim to provide more details on how these approaches have been applied to predict pathway memberships for genes with examples and caveats. We first discuss the different approaches based on the specific types of omics input data (Figure 1). Then we discuss how these heterogeneous input data can be integrated to predict metabolic pathways or networks using machine learning and offer a perspective on the future direction.
Assuming genes with high degrees of sequence and/or structural similarities tend to have similar functions, an early approach for predicting metabolic pathway membership is through evolutionary conservation. This approach is integral to the establishment of major metabolic pathway databases, such as Kyoto Encyclopedia of Genes and Genomes [28], MetaCyc [29] and Plant Metabolic Network [30]. These databases first ascribe the metabolic reaction of a gene product by predicting its Enzyme Commission numbers based on its protein sequence similarity to or the orthologous relationship with genes of known function. Then the target gene is assigned to metabolic pathways based on knowledge of the reactions catalyzed by pathway members. This approach can be error-prone because orthologs may not carry out biologically equivalent functions in different organisms [31]. In addition, minor changes in protein sequences can lead to altered reactions [32]. Furthermore, the sequences used in these databases may be incorrectly annotated based on sequence similarities, leading to errors in pathway membership predictions. Fortunately, these shortcomings can be alleviated by jointly considering protein domain signatures [33], molecular descriptors of gene sequences [34], K-mer representation of genes [35], and protein structures predicted with atomic accuracy [36].
Beyond sequences and structural similarity, the evolutionary histories of gene families across species have also been used for pathway prediction. This is based on the assumption that genes involved in the same pathway would be inherited together through speciation and should have similar phylogenetic profiles [37]. Another assumption is that, for a given pair of genes in species A, if their orthologs are functionally associated in species B, these two genes from A are likely to be functionally associated as well [38]. However, this approach can have a high false-positive rate because two genes in different pathways can have similar phylogenetic profiles by chance. It also has a high false-negative rate because the metabolite compositions may vary greatly among species. Thus, gene pairs catalyzing unique reactions will be overlooked.
Since one of the first plant MGCs was reported that contains genes of enzymes responsible for the biosynthesis of cyclic hydroxamic acid DIBOA in maize [39], >30 additional MGCs have been identified in plants [19,40e42]. The most straightforward approach to identify MGCs is by specifying a threshold number of Types of input data used in the computational prediction of plant metabolic pathway memberships. Arrows indicate the type of data used in different analyses that aid in the pathway membership predictions. EC: Enzyme Commission numbers; GO: Gene Ontology; GWAS: Genome-Wide Association Study.
enzyme genes that need to be in close proximity in the genomes. But this approach would have a high falsepositive rate. For example, on chromosome 7 in cultivated tomato, there are 10 genes belonging to acyl-CoA synthetase, BAHD acyltransferase, and enoyl-CoA hydratase families, but only three of them were specifically expressed in trichomes and were experimentally validated to be responsible for the medium-chain acylsugar biosynthesis in trichome [41]. This issue is addressed by incorporating additional evidence beyond genomic proximity with computational tools plantiS-MASH, PhytoClust, and PlantClusterFinder [30,43,44].
PlantiSMASH requires the enzyme genes in an MGC to contain 2 different protein domains [43], while PlantClusterFinder requires clustered genes encoding enzymes that catalyze 2 reactions [30]. In addition, all three tools allow consideration of gene co-expression which further improves prediction accuracy. Such approaches incorporating >1 evidence improve prediction accuracies. Joint consideration of additional features of pathway genes, such as genetic interactions or proteine protein interactions (discussed in a later section), can potentially further improve MGC predictions.
Genes in the same metabolic pathway have tightly associated functions, and are more likely to be coregulated and expressed in similar manners. Thus, a gene A that shares similar expression profiles with gene B in a known metabolic pathway may belong to the same pathway [41,45e48]. Non-enzymatic genes required for metabolite biosynthesis, such as transporters and transcriptional regulators, can also be revealed by coexpression analysis (reviewed in the study by Mutwil [26]). There are typically three ways to perform coexpression analysis (reviewed in the study by Delli-Ponti et al. [49]): (1) selecting the top most genes based on expression similarity; (2) clustering of expression profiles; (3) establishing co-expression networks. However, not all the genes within the same pathways have similar transcript profiles [16,17,49]. Among the 12 higher-order metabolic classes (each class contains a collection of interconnected pathways) investigated in the study by Wisecaver et al. [50], only genes involved in the biosynthesis of 'secondary metabolites' and 'cell structures' were significantly enriched in the coexpression network modules. In addition, the degree of gene co-expression within pathways is affected by expression datasets, expression values, and similarity measures [16,17]. For example, genes in the fatty acid aoxidation I and chlorogenic acid biosynthesis II pathways only showed high expression similarities when combined data of stress treatment was used, consistent with the role of these pathways in protecting plants against environmental perturbations [17].
It is also shown that different pathways differ widely in how good the predictions are after exhaustively considering how expression data can be used for predictions [17]. One approach to improve co-expression analysis is to consider protein levels in addition to transcript profiles. When proteomics data were used to identify genes co-expressed with Sterol side chain reductase 2 gene, the number of candidate genes for the cholesterol biosynthesis pathway dropped from 75 (based on transcript coexpression) to 33, narrowing down the candidates to be experimentally validated [46].
Beyond co-expression, other pre-established molecular networks, such as those based on genetic interactions and proteineprotein interactions, can be used to predict pathway memberships. Inferring genetic interactions, through analyzing single and double mutant phenotypes, is a classical approach for revealing function associations between genes in pathways [51]. Although w3.5% of current experiment-based pathway annotations are derived from genetic interaction data in Arabidopsis [52], such data have not been applied for predicting the pathway memberships of unknown enzyme genes.
Similar to genetic interactions, proteineprotein interactions have also been used widely to reveal functional association between genes within a pathway [53]. In addition, proteins catalyzing successive enzymatic steps often interact physically to form multienzymes complexes [54]. Proteineprotein interactions have also been combined with chemicale chemical and chemicaleprotein interactions to assign genes to metabolic pathways in yeast [55]. Including proteineprotein interaction data, AraNet and RiceNet established gene networks with co-citation, coexpression, domain co-occurrence, genomic neighborhood, and phylogenetic profile [38,56]. Because these gene networks contain functional association inference genome-wide, they can be used directly to hypothesize the metabolic pathway memberships of unknown enzyme genes based on their functional association to genes in known pathways.
Beyond networks containing geneegene or proteine protein relationships, metabolite networks have also been established by clustering mass spectrometry spectra [53] or incorporating additional features such as ion mobility separation with Feature-Based Molecular Networking [54]. Such metabolite networks can be incorporated into molecular networks connecting metabolic enzymes to substrates and/or products. Alternatively, the features for building metabolite networks can be jointly analyzed with features derived from other omics data to establish an integrated network containing both enzyme genes and associated metabolites.
Beyond the 'guilt-by-association' approaches discussed so far, another way to link genes to metabolic pathways is through metabolic GWAS (mGWAS) and metabolic quantitative trait locus analysis (mQTL) [45,57e63]. The rationale is that, by identifying the genetic variants that are associated with metabolic variants in a diversity panel or an inbred population, the genes where the associated genetic variants are located in or close to are likely members of the same pathway for the biosynthesis of specific metabolites. As a result, no prior information on pathway membership is needed and novel pathways can potentially be identified. mGWAS/mQTL analyses lead to thousands of genetic variants associated with a specific metabolic trait that may correspond to tens to hundreds of genes.
One challenge of the mGWAS/mQTL approach for predicting pathway genes is the difficulty in identifying the candidate genes that ultimately impact metabolic variation. Researchers either focus on the most significant variants [45] or do further data mining to narrow down the candidate causal genes [60]. For further data mining, the ProGeM framework [64], which is used in human disease studies, can potentially be repurposed for plant metabolic pathway predictions genome-wide. This framework prioritizes the candidate causal genes by leveraging the proximity of genes to the most significant trait-associated variants, an association of gene expression with the variants, and metabolic-related annotations from databases (i.e., Gene Ontology, Kyoto Encyclopedia of Genes and Genomes, Mouse Genome Informatics, Orphanet and Reactome) [64]. Another potential approach to identify candidate genes is based on the observation that the sequence variants associated with variation of a large set of metabolites are frequently located in hotspots [57,65,66]. These hotspots tend to be enriched in genes involved in the metabolic process and catalytic activity compared with background [57] or overlap with regions subject to strong positive selection (selection sweeps) [65,66]. Furthermore, for the majority of metabolites that are associated with more genes than expected, the associated genes tend to be clustered in the genome [65]. These findings highlight the possibility of jointly considering mGWAS/mQTL hotspots, local gene clusters, and strength of selection for pinpoint causal genes as well as those belonging to the same pathways.
In previous sections, we discussed approaches using distinct datasets for predicting metabolic pathway genes. In some examples, more than one type of data are used to improve prediction. For example, AraNet jointly considers multiple molecular networks [38]. As more data become available for plant species, particularly those generated with high-throughput omics technologies, the limitation of single data, for example, false positives in MGC detection, co-expression networks and mGWAS analysis, can be partly overcome by incorporating additional datasets [67,68]. For example, through the integration of phylogenetic, metabolomic and transcriptomic analysis, genes responsible for 4hydroxyindole-3-carbonyl nitrile [69], flavonoid [70], protolimonoid [71], and colchicine [72] biosynthesis pathways have been identified. In these cases, the candidate genes were first identified using one type of data, and then progressively filtered with other data. Another example is the establishment of a multi-omics network consisting of 13,361 metabolite-SNP-gene relationships by combining mGWAS, eQTL and metaboliteetranscript correlation analysis [66]. Using this network, candidate genes contributing to the flavonoids and steroidal glycoalkaloid biosynthesis pathways were hypothesized and tested [66].
While these studies are successful in identifying genes responsible for specific metabolic pathways and metabolites, a similar approach has yet to be applied to the genome-wide prediction of metabolic pathways. In addition, data beyond mGWAS, eQTL, and metabolomics are yet to be incorporated, such as comparative genomics, gene clustering, and co-expression among others. Through the integration of sequence similarities, genomic context, the likelihood of interactions between molecules, chemoinformatics, and ligand-binding experiments, Calhoun et al. [73] (2018) developed a computational method to place genes into bacterial metabolic pathways. This approach is only able to predict a gene into a linear metabolic pathway and has not been evaluated in any plant species.
The integration of multi-omics data has the potential but is rarely carried out for predicting metabolic pathways genome-wide in plants. To integrate the large and heterogeneous data for such a purpose, machine learning approaches are particularly suitable because they excel at learning patterns from highly complex datasets to make predictions [74]. Thus, there is an opportunity to use machine learning to predict plant metabolic pathway memberships. For the machine learning approaches to work well, there are four considerations.
First, a well labeled and abundant dataset, that is, highquality pathway annotations of enzyme genes, is crucial. Here, the label (e.g., a gene is a member of pathway X) is what we intend to predict. The challenge is that there are only a few plant species with relatively abundant experimentally validated pathway annotations [26]. Even if there are abundant annotations, such as in Arabidopsis, there are less than six genes annotated in a pathway on average. The challenge is that typically tens to hundreds of examples will be needed for training a good model. To alleviate this, pathway annotations from multiple species can potentially be integrated together for model training. Alternatively, genes can be first assigned to superpathways d a group of individual pathways which are associated with each other, then be predicted to individual pathways within each superpathway. In species with little or no experimentally derived pathway information, predictive models trained in model species can be applied to those data-poor species via transfer learning, a suite of machine learning algorithms for transferring knowledge from one domain to the other [75]. An example is a study of specialized metabolism gene predictions in tomatoes leveraging information from Arabidopsis [76].
Second, machine learning algorithms learn how to predict labels (i.e., pathways in this context) using a range of predictive variables referred to as features. Example features can include sequence similarity, genomic proximity, co-expression, and interaction information (Figure 1).
The success of some of the approaches discussed in the previous sections, such as AraNet, lies in the integration of multiple types of data. Although multi-omics data is rapidly accumulating, the data are heterogeneous in the data type, quality, genetic backgrounds, and experimental conditions for answering vastly different biological questions. There are also new data types to consider such as single-cell omics techniques, including transcriptome [77], methylome [78], and potentially metabolome [79] and proteome [80]. There is a significant effort required for data collection, cleaning, and transformation prior to machine learning modeling. In addition, because there can be a very large number of features, frequently it will be important to select features likely informative prior to building the prediction model using feature selection methods or to merge features with dimension reduction techniques [81].
Third, there is a wide range of machine learning frameworks and algorithms, and there is no guarantee which one is the best. For example, pathway prediction can be carried out by contrasting one pathway against the other (binary classification), by predicting a gene as belonging to one of many pathways at once (multi-class Prediction of plant metabolic pathway memberships using machine learning-based approaches. To integrate multi-omics data for predicting plant metabolic pathway memberships, two strategies are discussed here. One strategy (blue arrows) is to train a classical machine learning model by combining the features (predictive variables) obtained from each omics data. The other strategy is that, for each omics data, first establish the adjacency matrix among genes and make predictions using Graph Convolutional Networks (GCN), and then these initial omics-specific predictions can be used to make the final prediction using View Correlation Discovery Network (VCDN) [85]. Red and gray colored cells in 'Omics-specific prediction' and 'Final prediction' indicate that a gene (y-axis) is predicted as belonging to a pathway (x-axis). classification), or by predicting the probabilities of a gene as belonging to many pathways (multi-label classification) [82]. There is also an abundance of algorithms, including classical algorithms such as Random Forest [83] and Support Vector Machines [84], as well as more recent approaches such as Graph Convolutional Networks [85] and Transformer-based algorithms [87].
The key is to experiment with different frameworks and algorithms, as well as to ask which one performs the best and why.
Finally, multi-omics data can be integrated with machine learning to train pathway membership prediction models via two general strategies: direct feature integration and ensemble integration (Figure 2). In direct feature integration, features derived from different types of omics data are concatenated to build a combined feature matrix to train pathway prediction models. In the second strategy, features derived from each omics data type are used to build an omics-specific prediction model first. For each gene we want to predict the pathway membership, a prediction score is obtained using each omics-specific model. Then the prediction scores of all omics-specific models are used as features (the ensemble step) to build a final pathway prediction model. This strategy has been used to predict cancer types by integrating mRNA, miRNA, and methylation data [85].
Through such data integration exercises, models capable of predicting pathway memberships can be established and applied to predict genes of enzymes with no known pathway annotation that can be further tested experimentally. Beyond making predictions, the models themselves are also valuable for understanding why genes may belong to certain pathways. Such understanding can be obtained through the application of model interpretation methods where important features and their combinations that are important for the underlying prediction can be identified [86].
This study investigated the versatility of gene expression data for predicting plant metabolic pathway membership, by utilizing 656 combinations of expression datasets, expression values, and similarity measures. The authors concluded that unsupervised and supervised machine learning approaches outperformed the naive prediction (solely based on gene-to-pathway expression similarity) by a large margin. The results highlighted the need to extensively explore the expressionbased features and machine learning prediction strategies to optimize the pathway membership prediction of enzyme genes.
. Baranwal M, Magner A, Elvati P, Saldinger J, Violi A, Hero AO: A deep learning architecture for metabolic pathway prediction. Bioinformatics 2020, 36:2547-2553. This study adopted graph convolutional networks to extract molecular shape features of a biochemical compound, and then fed these features to a random forest classifier to predict classes of metabolic pathways in which the given compound participates.
. Mutwil M: Computational approaches to unravel the pathways and evolution of specialized metabolism. Curr Opin Plant Biol 2020, 55:38-46. Reviewed the evolution and diversity of metabolic pathways in the Archaeplastida, and discussed computational approaches to identify specialized metabolic pathways, including approaches based on sequence similarity, biosynthetic gene clusters and co-expression analysis.
. Jacobowitz JR, Weng J-K: Exploring uncharted territories of plant specialized metabolism in the postgenomic era. Annu Rev Plant Biol 2020, 71:631-658. Reviewed recent progresses in de novo biosynthesis pathway discovery, which employed biochemical or computational approaches. Also highlighted the higher-order organization of plant specialized metabolism.
Physiology and metabolism (2022) Current Opinion in Plant Biology 2022, 66:102171 www.sciencedirect.com
Physiology and metabolism (2022) Current Opinion in Plant Biology 2022, 66:102171 www.sciencedirect.com
Current Opinion in Plant Biology 2022, 66:102171 www.sciencedirect.com