Skip to main content
NCBI home page
G3: Genes | Genomes | Genetics logoLink to G3: Genes | Genomes | Genetics
. 2022 Sep 23;12(11):jkac255. doi: 10.1093/g3journal/jkac255

Prioritizing candidate eQTL causal genes in Arabidopsis using RANDOM FORESTS

Margi Hartanto 1,✉, Asif Ahmed Sami 2, Dick de Ridder 3, Harm Nijveen 4,✉
Editor: J Ross-Ibarra
PMCID: PMC9635658  PMID: 36149290

Abstract

Expression quantitative trait locus mapping has been widely used to study the genetic regulation of gene expression in Arabidopsis thaliana. As a result, a large amount of expression quantitative trait locus data has been generated for this model plant; however, only a few causal expression quantitative trait locus genes have been identified, and experimental validation is costly and laborious. A prioritization method could help speed up the identification of causal expression quantitative trait locus genes. This study extends the machine-learning-based QTG-Finder2 method for prioritizing candidate causal genes in phenotype quantitative trait loci to be used for expression quantitative trait loci by adding gene structure, protein interaction, and gene expression. Independent validation shows that the new algorithm can prioritize 16 out of 25 potential expression quantitative trait locus causal genes within the top 20% rank. Several new features are important in prioritizing causal expression quantitative trait locus genes, including the number of protein–protein interactions, unique domains, and introns. Overall, this study provides a foundation for developing computational methods to prioritize candidate expression quantitative trait locus causal genes. The prediction of all genes is available in the AraQTL workbench (https://www.bioinformatics.nl/AraQTL/) to support the identification of gene expression regulators in Arabidopsis.

Keywords: Arabidopsis thaliana, eQTL, gene expression, machine learning, causal gene

Introduction

One of the main objectives of genetic research is to link traits to genotypic variation. However, the path from genetics to observable traits is not straightforward; instead, it goes through a network of interconnecting intermediate phenotypes, such as gene expression, protein levels, and metabolite levels (Civelek and Lusis 2013). Studying the effect of genetic perturbation on these intermediate phenotypes could improve our understanding of how a trait is regulated. Following recent advances in omics technology, the effect of multiple genetic perturbations can now be studied in a single experiment using linkage mapping or association studies. One example is genetical genomics, where variation in transcript levels is statistically associated with genetic variation in a population (Jansen and Nap 2001) to find so-called expression quantitative trait loci (eQTLs).

A mapped eQTL can be categorized as cis or trans based on its location relative to the affected gene. Cis-eQTLs are mapped close to the gene and are assumed to arise due to sequence polymorphisms in or near the gene itself, for instance, in cis-regulatory elements (e.g. the promoter). In contrast, trans-eQTLs are mapped far away from the target gene and emerge due to polymorphisms in trans-acting factors (e.g. transcription factors) called expression quantitative trait genes or eQTGs (Brem et al. 2002; Rockman and Kruglyak 2006). However, a trans-eQTL typically spans a large genomic region with hundreds of candidate eQTGs. Experimental fine mapping to narrow down the region (e.g. in Eshed and Zamir 1995) is costly and laborious. As a result, only a few causal genes have been identified in the thousands of eQTLs that have been mapped for Arabidopsis thaliana, using different populations and experimental conditions (Keurentjes et al. 2007; West et al. 2007; Cubillos et al. 2012; Snoek et al. 2012; Lowry et al. 2013; Hartanto et al. 2020). As an in silico alternative, a prioritization method can help to limit the number of candidate eQTGs for further validation.

Several network-based methods have been used to find eQTGs (e.g. in Keurentjes et al. 2007; Jimenez-Gomez et al. 2010; Hartanto et al. 2020). These methods primarily aim to find master regulator(s) at loci where trans-eQTLs for many genes are collocated, known as eQTL hotspots (Breitling et al. 2008). In general, these methods utilize a coexpression network built using genes having an eQTL on the hotspot (called targets) and genes located in the hotspot (called candidate eQTGs). Candidates are then usually prioritized based on a network centrality measure, such as degree centrality (i.e. the number of genes interacting with a candidate) or closeness centrality (i.e. the average path length between a candidate and all other genes) (Serin et al. 2016; Hartanto et al. 2020). Several candidate eQTGs have been identified in this way, for example, GIGANTEA (Keurentjes et al. 2007), ELF3 (Jimenez-Gomez et al. 2010), ICE1, and DEWAX (Hartanto et al. 2020). This approach, unfortunately, only works for eQTL hotspots, not for regions that only have a small number of eQTLs. Another limitation is the sole reliance on coexpression data: given the complexity of gene expression regulation, the expression of the regulator is not necessarily correlated to that of its targets, particularly in eukaryotes (Lelli et al. 2012; Marbach et al. 2012). Therefore, additional data sources should be considered to capture possible interactions between the regulator and its target.

Previously, a machine-learning-based method, QTG-Finder, was developed to prioritize candidate genes for phenotype QTLs in Arabidopsis (Lin et al. 2019). This method used features derived from various gene properties, such as paralog copy number, gene ontology (GO), and the number of SNPs, to rank the candidate genes in the QTL interval. The model could recall 64% of Arabidopsis QTGs when the top 20% ranked genes were considered. Further development of this method led to QTG-Finder2, which used orthology information and allowed for gene prioritization in species with no or few known QTGs (Lin et al. 2020). We were curious about the capability of this algorithm to prioritize eQTGs, given that some QTGs are involved in gene expression regulation, for example, ELF3 (Jimenez-Gomez et al. 2010), ERECTA (Terpstra et al. 2010), FRI (Lowry et al. 2013), MAM1 (Jansen et al. 2009), and AOP2 (Jansen et al. 2009).

We propose eQTG-Finder, an extended version of QTG-Finder2 for eQTG prioritization, and apply the new algorithm to prioritize eQTGs in Arabidopsis. eQTG-Finder contains 12 new features based on protein–protein interaction (PPI), gene structure, and expression variation. Three of these features significantly improve model performance, which is underscored by a feature importance analysis. We demonstrate the efficacy of this algorithm in prioritizing eQTGs using an independent test set. Finally, we use the new model to predict all Arabidopsis genes and make these available in our Arabidopsis eQTL analysis platform AraQTL (https://www.bioinformatics.nl/AraQTL/)(Nijveen et al. 2017) to help identify gene expression regulators.

Materials and methods

QTG-Finder2 was developed for prioritizing causal phenotype QTL genes (QTGs) in Arabidopsis (Lin et al. 2020). This algorithm consists of 5,000 Random Forest classifiers (Ho 1998) trained using known QTGs and Arabidopsis orthologs of QTGs from other species as positives and other genes as negatives. QTG-Finder2 prioritizes candidate genes based on features generated from polymorphism data, functional annotation, cofunction networks, and paralog copy numbers. Our method extends QTG-Finder2 with new features, and we train the resulting model using the same sets of positive and negative genes. We evaluate the performance in prioritizing candidate causal eQTL genes (eQTGs) in Arabidopsis.

New features

We generate and include 12 new features in addition to the ones already used by QTG-Finder2. These new features are based on PPI, gene expression, and gene/protein structure.

PPI feature

Genes can be associated with other genes, for instance, because the encoded proteins participate in the same pathway or are mentioned in the same publication. The number of such interactions a gene has could measure its propensity to be an eQTL causal gene. We generate a network-based feature using Arabidopsis PPI data from STRING-DB (Szklarczyk et al. 2019). The data were downloaded from the download page of STRING-DB version 11 (https://string-db.org/cgi/download). We only keep high-confident interactions by removing those with STRING scores below 700. We count the number of interactions of each Arabidopsis gene as a feature.

Gene expression features

We previously showed that different stages of seed germination each have a unique eQTL landscape pointing to stage-specific regulators (Hartanto et al. 2020). This indicates that variation in gene expression may help distinguish eQTL causal genes from other (noncausal) genes. We, therefore, generate 7 features based on the average and standard deviation of gene expression across different tissues, accessions, and conditions (control vs. treatments):

Tissues

We downloaded RNA-seq data for 9 different tissues (flower, root, male organ, seeds, female organ, stem, leaf, apical meristem, and root meristem) from CoNekT (http://www.evorepro.plant.tools/) (Julca et al. 2020). For each gene, the standard deviation is calculated and used as a feature (“SD exp. Across tissues”).

Accessions

We used RNA-seq data measured in seedlings of 19 different Arabidopsis accessions (Zu-0, Wu-0, Ws-0, Wil-2, Tsu-0, Sf-2, Rsch-4, Po-0, Oy-0, No-0, Mt-0, Ler-0, Kn-0, Hi-0, Edi-0, Ct-1, Col-0, Can-0, and Bur-0). These data are obtained from the Arabidopsis RNA-seq Database (http://ipf.sustech.edu.cn/pub/athrna/) (Zhang et al. 2020). The average and standard deviation were calculated and used as features (“avg exp. across accessions” and “SD exp. across accessions”).

Conditions

From the same database, we collected whole tissue RNA-seq data of the wild-type Col-0 accession. We divided these data into experiments with and without treatments to generate 4 features for average and standard deviation of treatment and control conditions (“avg exp across treatments,” “avg exp. across controls,” “SD exp. across treatments,” and “SD exp. across controls”).

We removed datasets from the Arabidopsis RNA-seq Database with a very low total read count and/or many unmapped reads. The list of samples used to generate gene expression features can be found in Supplementary Table 6.

Structural features

The structure of causal genes and encoded proteins might differ from the other genes. Therefore, we generate 4 structural features: the numbers of introns, total protein domains, unique protein domains, and splice variants per gene. Data were retrieved from https://www.arabidopsis.org/ (accessed May 2021). The number of introns and splice variants are counted in TAIR10's BLAST datasets. The other 2 features are generated from all.domains.txt by counting each Arabidopsis gene's total number of domains and the number of unique domains.

Hyperparameter tuning

Model evaluation is based on QTG-Finder (Lin et al. 2019) and QTG-Finder2 (Lin et al. 2020). Given the low number of known eQTGs, we use known QTGs and Arabidopsis orthologs of QTGs found in other species as positives and other genes as negatives, similar to QTG-Finder2. We use hyperparameter tuning to determine the best parameter combination (the number of trees, minimal samples split, and maximum number of features) using grid search and assess the area under the curve (AUC) of the receiver-operating characteristic (ROC) curve in an extended version of the 5-fold cross-validation framework. In this framework, the positives are randomly re-split into a training and validation set in a 4:1 ratio iteratively. Next, each set is combined with randomly selected negatives. The ratio of positives and negatives is an optimized hyperparameter. This splitting of positives is done 50 times, and for each positive set random selection of the negatives was conducted 50 times. This extensive procedure (2,500 evaluations) makes that positive cooccurs with all negative at least once with high probability. All machine-learning model training and testing in this study is performed using Python’s scikit-learn library version 1.0.2.

Selection of candidate eQTL genes and independent validation of model performance

A list of candidate eQTGs in Arabidopsis is manually selected from the literature. These genes are categorized as confirmed/strong-candidate, hypothetical, or hypothetical-ortholog. Genes that have been through experimental validation or have strong evidence as eQTG are categorized into the confirmed/strong-candidate group, for example, GIGANTEA (Keurentjes et al. 2007; Snoek et al. 2012). Some confirmed/strong-candidate eQTGs are used as positive in QTG-Finder2, and we remove these from the positive instances to be used as validation genes. Meanwhile, genes that were not experimentally validated but are predicted to play a role as eQTG through in silico analysis (e.g. network analysis) are categorized as hypothetical, for example, ICE1 and DEWAX (Hartanto et al. 2020). If a gene's ortholog is considered an eQTG in another species, it is categorized as hypothetical-ortholog; for example, NF-YC4 is found as an eQTG in potatoes (van Muijen et al. 2016). In total, this yields 25 candidate eQTGs in Arabidopsis: 6 confirmed/strong-candidate, 4 hypothetical, and 15 hypothetical-ortholog genes (Supplementary Table 1). We ensure that these candidates are not used for hyperparameter tuning or cross-validation.

Independent validation is performed using the best combination of parameters (Supplementary Table 5). We train 5,000 Random Forest classifiers using all positives but different sets of negatives, with a positive: negative ratio of 1:200 to approximate the ratio of causal and noncausal genes in real eQTLs. The models are then applied to each candidate eQTG and other genes located within 2 Mbp around it (1-Mbp upstream and 1-Mbp downstream). For these genes, the average probability of being causal is calculated over 5,000 models. These average probabilities are then ranked for prioritization, and the rank is calculated as a performance measure. For example, a rank of 10% indicates that 10% of genes in the eQTL region rank higher than the candidate.

Feature importance analysis

Feature importance is determined using a leave-one-out analysis. Iteratively, each feature is removed from the dataset, and a model is trained using the reduced dataset. The AUC difference between the full model (with all features) and the reduced model is then calculated and used to indicate the feature importance. In addition, we calculate feature importance for clusters of correlated features. Features are clustered if they have a pairwise Pearson correlation equal to or larger than 0.6. We use the previous cross-validation framework and the best parameters to measure the model performance in this analysis.

Data analyses

Pairwise Pearson correlation coefficients between features are calculated using the Pandas (version 1.3.5) DataFrame.corr method in Python. Pearson Wilcoxon rank sum test analyzes differences in the median between positive and negative genes for the 12 new features. The test is conducted in R using the base “wilcox.test” function. GO enrichment analysis for the top and bottom 5% predicted causal genes is performed using TopGO in R (Alexa et al. 2006) using the algorithm's default “weight01” parameter, which is the mixture of “elim” and “weight” methods. The Python version used for the analyses is 3.8.12, and the R version is 4.0.2.

Results

The QTG-Finder2 algorithm could rank phenotype QTL causal genes higher than other genes in a cross-validation setting (AUC = 0.81) and recall 80% independent curated causal genes when the top 20% of genes in the QTL are considered (Lin et al. 2020). In this study, we extend QTG-Finder2 with a set of new features and evaluate its performance in prioritizing expression QTGs.

New features improve causal gene prediction performance

To improve model performance and better tailor it fit for eQTG prioritization, we added 12 new features based on gene expression, structure, and PPI in the QTG-Finder2 algorithm. Most new features only show a low to moderate correlation with the existing ones (Supplementary Fig. 1), indicating that we add new information to the model. Figure 1 shows feature distributions for the causal genes as the positive class (55 known QTGs and 145 Arabidopsis ortholog of QTGs from other species) and the other genes in the genome as the negative class (n = 26,970). For most features, the causal genes' median value is significantly different from that of the other genes in the genome (see Supplementary Table 2). The expression of causal genes is more variable than that of other genes. Moreover, causal genes tend to have more and varied protein domains. Causal genes also have slightly more introns than other genes. These differences between the causal genes and the other genes in the genome provide a first indication of potential discriminating features for the machine-learning model. We assess the performance of the model with and without new features using a cross-validation framework.

Fig. 1.

Fig. 1.

Open in a new tab

Distribution of 12 new features for known causal genes as the positive class (n = 200; 55 known QTGs and 145 orthologs of QTGs from other species) and the remaining genes in the genome as the negative class (n = 26,970). Significance of differences in medians was assessed using the Wilcoxon rank sum test (*P ≤ 0.05; ****P ≤ 0.0001). Red dots indicate means. SD, standard deviation; Exp., gene expression.

To assess the contribution of new features to the model performance, we compare the AUC of the ROC between the original QTG-Finder2 and the extended model that we labeled eQTG-Finder, and for the extended model with the class labels permutated, as a control (Fig. 2a). The AUC was measured in an extended cross-validation setting over 2,500 different combinations of positive and negative gene sets. The results show that eQTG-Finder (AUC = 0.859 ± 0.008) performs better than QTG-Finder2 (AUC = 0.801 ± 0.01) and the control model (AUC = 0.502 ± 0.014). Adding new features thus allows the model to rank causal genes higher than the other genes. The next section analyzed model performance in prioritizing eQTG using selected candidate eQTGs.

Fig. 2.

Fig. 2.

Open in a new tab

a) AUC of the ROC of the original QTG-Finder2 model (blue) and extended eQTG-Finder model (green), and eQTG-Finder trained with randomized class labels (red) as a control. Transparent areas indicate standard deviations over 2,500 repetitions. b) Feature importance is measured using leave-one-out analysis. A positive AUC drop indicates that the removal of the feature reduces the model's predictive capability. Feature names in bold and with dark blue bars indicate new features. Error bars indicate standard deviations over 2,500 repetitions.

To determine how the new features contribute to causal gene prediction, we calculate feature importance using a leave-one-out approach (Fig. 2b). Each feature is iteratively removed from the dataset, and the reduced model’s performance is compared to that of the model containing all features. The drop in AUC indicates a feature's importance. A positive AUC drop means removing that feature decreases the model's predictive capability. The result shows that 4 of the most important features in the model are the new ones: the number of unique domains, the PPI count, the intron count, and the domain count. However, the large standard deviation for the domain count AUC drop indicates that the contribution of this feature is not consistent over different samples of positive and negative sets.

Some features in the model are highly correlated (Supplementary Fig. 1). When one of these features is removed to calculate feature importance, the reduced model will resort to using these correlated features. As a result, the removed feature might be assigned lower importance than it should have in the model (Gregorutti et al. 2017). To avoid this, we calculated feature importance for clusters of features. The result (Supplementary Fig. 3) shows a slight change in the importance of some features, for example, “network weight” is now among the top important features since it is correlated with “ppi count.”

eQTG-Finder ranks most strong eQTG candidates better than QTG-Finder2

To evaluate eQTG prioritization performance, we again train the original QTG-Finder2 and the extended eQTG-Finder model and use them to rank selected potential eQTGs (Supplementary Table 1). Models are trained using all positives (known QTGs and Arabidopsis ortholog QTGs from other species). We repeated the training 5,000 times with different negative samples to select each negative gene at least once in training with high probability. These models rank each of the 25 potential eQTGs with their surrounding genes within a 2-Mbp window as a hypothetical eQTL region. These potential eQTGs are selected manually from the literature and grouped based on the evidence of being causal eQTL genes (see Materials and Methods for detail). Gene ranking is based on the average probability of a gene being causal, as predicted by the 5,000 models. We use the rank to indicate the percentage of genes on the eQTL with higher ranks than the gene of interest (i.e. a rank of 10% indicates that 10% of genes in the eQTL region rank higher than the gene of interest). We predefine cutoffs of 5%, 10%, and 20%, in each of which we compare recall between QTG-Finder2 and eQTG-Finder. These recalls for different cutoffs can be used by researchers to decide the proportion of top prioritized genes for further experimental validation.

The QTG-Finder2 model recalls 16%, 28%, and 52% of eQTG candidates if the top 5%, 10%, and 20% ranked genes are considered (Fig. 3). With added features, eQTG-Finder ranks eQTGs slightly better with percentages of 36%, 52%, and 64% respectively. The eQTGs vary in their evidence of being causal genes (see Materials and Methods). Four out of 6 strong eQTG candidates (AOP2, ERECTA, GIGANTEA, and MAM1) rank within the top 5% by eQTG-Finder compared to only one (ERECTA) by QTG-Finder2. The other 2 strong candidates, FRI and ELF3, were ranked at 10.2% and 61.2% by eQTG-Finder. The ranks of 16 genes are improved by eQTG-Finder, 8 are worse, and 1 stays the same (Supplementary Table 3). The rank of 4 out of 6 strong eQTG candidates improves, with GIGANTEA one of the most drastic improvements, moving from 53.7% to 4.2%. On the other hand, the rank of ERECTA drops (0.4–2.8%) but remains in the top 5%. Both models rank another strong eQTG candidate ELF3 poorly (at 44% by QTG-Finder2 and 61.2% by eQTG-Finder). As the number of strong eQTG candidates is limited, we also show the prioritization of hypothetical and hypothetical orthologs eQTGs. Even though the improvement was not as large as for the strong eQTG candidates, eQTG-Finder still ranks most of the hypothetical and hypothetical-ortholog eQTGs in the top 10%.

Fig. 3.

Fig. 3.

Open in a new tab

Rank comparison of 16 candidate eQTGs using the model with new features (eQTG-Finder) and the original model (QTG-Finder2).

Despite the decent overall performance in candidate eQTGs prioritization, we notice that eQTG-Finder performance in prioritizing phenotype QTGs is still inconsistent. Using the initial independent validation set, only 7 out of 11 QTGs are ranked within the top 20% by eQTG-Finder, compared to 9 by QTG-Finder2 (Supplementary Fig. 2).

To get an overview of eQTG-Finder predictions, we inspect the distribution of the average predicted probability of being causal for all Arabidopsis genes (Fig. 4). This skewed toward a low value, with a median value of 0.007 (note that the x-axis of Fig. 4 is on a log10 scale). Twenty-one of the 25 genes in the validation set have a predicted probability higher than the median. ELF3 (probability = 0.0045) is the only strong eQTG candidate with a predicted probability lower than the median.

Fig. 4.

Fig. 4.

Open in a new tab

The density plot of probabilities of being causal predicted by eQTG-Finder for all Arabidopsis genes. Text labels point to the probability of the gene in the plot. The x-axis is on a log10 scale.

A GO enrichment analysis shows that the top 5% genes in the distribution are significantly enriched (false discovery rate P-value <0.05) for 67 GO terms (Supplementary Table 4), most of which are related to response to abiotic and biotic stresses, such as “defense response to bacterium,” “defense response to fungus,” and “response to wounding.” The term “regulation of transcription” is also enriched, suggesting that transcription factors are likely to be causal, consistent with the feature importance analysis result where “is_TF” is among the most important features. Meanwhile, the bottom 5% are not enriched for any term.

eQTG-Finder is available in AraQTL to support new hypotheses on the gene expression regulation

To make eQTG-Finder results easily accessible for researchers, we include predicted probabilities of causality (herewith referred to as eQTG-Finder score) for all Arabidopsis genes in AraQTL, our Arabidopsis eQTL data workbench (Nijveen et al. 2017). Prioritizing genes using QTG-Finder2 is not straightforward as it requires users to prepare a list of candidate genes and command-line usage skills. Integrating the eQTG-Finder score in AraQTL facilitates users to interactively identify gene expression regulators. For example, we here discuss a case on predicting a new potential regulator for GLK2 using the eQTG-Finder score and other interaction evidence in AraQTL. GLK2 is a GARP nuclear transcription factor involved in light-controlled signaling (Waters et al. 2009). Liu et al. (2022) recently found that HY5 is the regulator of GLK2 based on the fact that HY5 is a well-known regulatory switch for light signaling in literature. The same conclusion can also be derived using the Serin et al. (manuscript in preparation) eQTL experiment and prior knowledge data in AraQTL. Another approach to finding potential regulators of GLK2 can be made in AraQTL using the eQTG-Finder score. In a Kas × Tsu eQTL experiment on leaf tissue (Lowry et al. 2013), GLK2 has an eQTL on the beginning of chromosome 1, indicating the location of the potential regulator(s) (Fig. 5, top). As many as 257 candidate regulatory genes are present in the eQTL (Fig. 5, bottom). We can filter out weak candidates by constructing a network of GLK2 connected to its potential regulators on the eQTL based on prior knowledge, such as PPI and gene annotation (Hartanto et al., manuscript in preparation). Here, we threshold the eQTG-Finder score to remove weak candidates. Moreover, eQTG-Finder can prioritize the remaining 14 genes by selecting the “Bipartite by eQTG-Finder score” network layout and ordering genes by their score. The result suggests some promising GLK2 regulator candidates ranked at the top, for example, a transcription factor LHY in second place. Until now, LHY has not been reported to regulate GLK2. However, this gene is a promising GLK2 regulator candidate as the network shows that it has a transcription factor binding site(s) on the GLK2 promoter (O'Malley et al. 2016). Moreover, LHY is involved in light signaling (Kim et al. 2003; Joo et al. 2017). This example suggests that integrating the eQTG-Finder score in AraQTL can help infer new regulatory interactions.

Fig. 5.

Fig. 5.

Open in a new tab

Prioritization of GLK2 regulator using the eQTG-Finder score in AraQTL. (Top) eQTL profile of GLK2 from the Lowry et al. (2013) experiment. The eQTL region on chromosome 1 (shaded in pink) pinpoints the location of potential GLK2 regulator(s). (Bottom) Prior knowledge network connecting GLK2 (blue node) with candidate regulators (yellow nodes) based on prior knowledge data. Here, the eQTG-Finder score is used to order candidates based on their probability of being causal.

Discussion

The concept of genetical genomics was first coined 2 decades ago (Jansen and Nap 2001), and numerous Arabidopsis eQTL data sets have been published since then (Nijveen et al. 2017). The aim of genetical genomics is to pinpoint genomic regions associated with gene expression variation (eQTL) and ultimately unravel genes involved in expression regulation. However, identifying causal genes (eQTGs) is difficult because of the often large genomic regions they span, regularly harboring dozens or even hundreds of candidates. The regions can be narrowed down by experimental fine-mapping (Eshed and Zamir 1995), and the remaining candidate genes can then be validated using functional genomics methods (e.g. using CRISPR-Cas9-mediated deletions as in Evans and Andersen 2020). However, performing these experiments for thousands of eQTLs is very costly. Using genomics and annotation data, a computational prioritization method can help identify candidate eQTGs. This study extends an existing machine-learning algorithm, QTG-Finder2, to address this issue and evaluates its performance in prioritizing eQTG. eQTG-Finder outperforms QTG-Finder2 in distinguishing positive causal genes from the other genes in the genome based on a cross-validation setting (Fig. 2a). Moreover, eQTG-Finder prioritizes most eQTGs in eQTLs better than QTG-Finder2 in an independent validation test (Fig. 3). We make eQTG-Finder scores available in AraQTL to help researchers interactively identify key regulators.

The key improvement of eQTG-Finder lies in the inclusion of 12 new features based on gene expression, structure, and interactions. Given the complexity of the resulting model, it is not straightforward to assess how these features improve eQTG-Finder in gene prioritization (Petch et al. 2022). We calculated the contribution of each feature in the model using a leave-one-out feature importance analysis (see Materials and Methods). This showed that the number of unique protein domains, the number of PPI and the number of introns are in the top 5 most contributing features in the model. We showed that known causal genes tend to have more domains, PPI partners, and introns than other genes (Fig. 1). These new features may provide insight into what distinguishes causal and noncausal genes. For instance, since protein domains determine protein functions (Enright and Ouzounis 2001; Vogel et al. 2004), the presence of multiple domains in a causal gene could indicate involvement in a wide range of biological functions. The diverse functions of causal genes could also be reflected in their larger number of PPI partners than noncausal as genes perform their function in concert with other genes (Ito et al. 2001). The number of introns reflects the number of exons in a gene. Several studies demonstrated that exons play a role in the evolution of domain architectures through exon-shuffling, leading to new combinations of domains with new functions.

Variation in phenotype can be traced back to variation in gene expression (Skelly et al. 2009; Albert and Kruglyak 2015). For this reason, we included features based on the standard deviation (SD) of gene expression across different Arabidopsis accessions, tissues, and conditions. Even though the medians between causal and other genes are significantly different (Fig. 1), features based on SD of expression have low importance in the model (Fig. 2a). A possible explanation for this could be that features based on expression are correlated (Supplementary Fig. 1) and, therefore, their importance is underestimated (Gregorutti et al. 2017). We, therefore, removed all of these correlated features and re-calculated the feature importance. The feature importance, however, remains the same. Nevertheless, we do not have evidence that these features negatively affect the prediction performance; hence, we kept them in the model.eQTG-Finder uses known QTGs (i.e. causal genes for a phenotype QTL) as positive instances for model training because of the limited number of known eQTGs. A recent finding in humans showed that cis-eQTLs and GWAS genes are different due to the detection bias of the assays (Mostafavi et al. 2022). This detection bias could also hold for trans-eQTL and phenotype QTL genes in Arabidopsis. However, we argue that QTGs are still relevant for prioritizing eQTG since variation at the molecular level (e.g. in gene expression, metabolite, or protein level) can be propagated and cause variation at higher phenotypic levels (Fu et al. 2009; Civelek and Lusis 2014). For example, genetic variations in AOP2 and MAM1 cause cis-eQTLs for gene expression and metabolite QTLs for aliphatic glucosinolate biosynthesis, which confer insect resistance in Arabidopsis (Wentzell et al. 2007; Jansen et al. 2009). Both genes were prioritized in the top 5% by eQTG-Finder. This result suggests that eQTG-Finder can identify QTLs for other molecular phenotypes, including metabolite and protein.

A lack of model interpretability may hamper a user’s comprehensive evaluation and assessment of the prioritization results. Regardless of the good performance, it is difficult to precisely understand how eQTG-Finder classifies certain genes as causal and others as noncausal, a typical issue for a complex model like Random Forest (Petch et al. 2022). Instead, in AraQTL, we provide additional sources of evidence to support the eQTG-Finder prioritization results (Hartanto et al., unpublished). For example, eQTG-Finder prioritizes transcription factor LHY as the regulator of GLK2 (Fig. 5). The network visualization in AraQTL showed that LHY is connected to GLK2 by transcription factor binding site evidence, indicating that LHY may bind to the GLK2 promoter and modulate its expression. Incorporating eQTG-Finder in the AraQTL web interface facilitates researchers to identify key regulators for genes of interest without the need for computational skills.

In the independent validation, some eQTG candidates were ranked poorly by eQTL-Finder (Fig. 3). Low-ranked assumed eQTG genes from the hypothetical and hypothetical-orthologs groups might not be actual eQTGs; however, the strong eQTG candidate ELF3 was also ranked poorly by both eQTG-Finder (61.2%) and QTG-Finder (44%). ELF3 encodes a nuclear protein and was demonstrated to regulate gene expression leading to shade-avoidance response (Jimenez-Gomez et al. 2010). The complexity of the eQTG-Finder algorithm makes it difficult to dissect the prediction for ELF3. We investigated 2 of the most important features and noticed that this gene only has 1 identified protein domain and 1 paralog copy number, which is lower than the median values of causal genes (4 and 17, respectively).

We observed that eQTG-Finder prioritization of candidate QTGs in independent validation was slightly worse compared to QTG-Finder2 (Supplementary Fig. 2), despite its better performance in cross-validation (Fig. 2b). The new expression-based features might bias eQTG-Finder toward prioritizing eQTGs compared to QTGs, but the complexity of the model makes it difficult to learn exactly how these features affect prioritization. Moreover, the number of 11 candidates we used for validation is too low to allow a very precise assessment of the general performance of eQTG-Finder in prioritizing QTGs.

Likely, some features associated with eQTG are still missing in our model or underrepresented in our set of positive instances. Since the regulator-target relationship is specific, we expect that features representing gene–gene/PPI [e.g. STRING scores (Szklarczyk et al. 2019), transcription factor binding sites (Tian et al. 2020), and GO semantic similarity (Yu 2020)] are relevant for prioritizing eQTG. Including these would shift the prioritization of generic eQTGs based on gene properties to the prioritization of eQTGs for a specific target using features based on gene-pair relationships. This is similar to the approaches of Wong et al. (2004) and Pandey et al. (2010), who predicted genetic interaction using gene pair relationships in yeast. The number of positive examples (i.e. confirmed eQTG-target pairs) is currently too small to properly train such a model for Arabidopsis. However, as data regarding genetic regulation is steadily increasing, we are optimistic that this strategy will be possible in the future.

Supplementary Material

jkac255_Supplementary_Table1_2_4_5
Click here for additional data file. (7.1KB, zip)
jkac255_Supplementary_Figure_S1
Click here for additional data file. (62.2KB, pdf)
jkac255_Supplementary_Figure_S2
Click here for additional data file. (390KB, jpeg)
jkac255_Supplementary_Figure_S3
Click here for additional data file. (249.7KB, jpeg)
jkac255_Supplementary_Table3
Click here for additional data file. (15.1KB, xlsx)
jkac255_Supplementary_Table6
Click here for additional data file. (16.6KB, xlsx)

Acknowledgments

The authors thank members of the Bioinformatics Group, Wageningen University, for feedback and suggestions.

Conflicts of interest

None declared.

Contributor Information

Margi Hartanto, Bioinformatics Group, Wageningen University and Research, 6708 PB Wageningen, The Netherlands.

Asif Ahmed Sami, Bioinformatics Group, Wageningen University and Research, 6708 PB Wageningen, The Netherlands.

Dick de Ridder, Bioinformatics Group, Wageningen University and Research, 6708 PB Wageningen, The Netherlands.

Harm Nijveen, Bioinformatics Group, Wageningen University and Research, 6708 PB Wageningen, The Netherlands.

Data Availability

The code and data for the analysis and visualization are available at the Wageningen University GitLab repository (https://git.wur.nl/harta003/eqtg-finder). eQTG-Finder prioritization is available at AraQTL (https://www.bioinformatics.nl/AraQTL/; Nijveen et al. 2017).

Supplemental material is available at G3 online.

Literature cited

  1. Albert FW, Kruglyak L.. The role of regulatory variation in complex traits and disease. Nat Rev Genet. 2015;16(4):197–212. [DOI] [PubMed] [Google Scholar]
  2. Alexa A, Rahnenführer J, Lengauer T.. Improved scoring of functional groups from gene expression data by decorrelating GO graph structure. Bioinformatics. 2006;22(13):1600–1607. [DOI] [PubMed] [Google Scholar]
  3. Breitling R, Li Y, Tesson BM, Fu J, Wu C, Wiltshire T, Gerrits A, Bystrykh LV, de Haan G, Su AI, et al. Genetical genomics: spotlight on QTL hotspots. PLoS Genet. 2008;4(10):e1000232. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Brem RB, Yvert G, Clinton R, Kruglyak L.. Genetic dissection of transcriptional regulation in budding yeast. Science. 2002;296(5568):752–755. [DOI] [PubMed] [Google Scholar]
  5. Civelek M, Lusis AJ.. Systems genetics approaches to understand complex traits. Nature Reviews Genetics. 2014;15(1):34–48. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Cubillos FA, Yansouni J, Khalili H, Balzergue S, Elftieh S, Martin-Magniette M-L, Serrand Y, Lepiniec L, Baud S, Dubreucq B, et al. Expression variation in connected recombinant populations of Arabidopsis thaliana highlights distinct transcriptome architectures. BMC Genomics. 2012;13:117. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Enright AJ, Ouzounis CA.. Functional associations of proteins in entire genomes by means of exhaustive detection of gene fusions. Genome Biol. 2001;2(9):RESEARCH0034. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Eshed Y, Zamir D.. An introgression line population of Lycopersicon pennellii in the cultivated tomato enables the identification and fine mapping of yield-associated QTL. Genetics. 1995;141(3):1147–1162. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Evans KS, Andersen EC.. The gene scb-1 underlies variation in Caenorhabditis elegans chemotherapeutic responses. G3 (Bethesda). 2020;10(7):2353–2364. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Fu J, Keurentjes JJB, Bouwmeester H, America T, Verstappen FWA, Ward JL, Beale MH, de Vos RCH, Dijkstra M, Scheltema RA, et al. System-wide molecular evidence for phenotypic buffering in Arabidopsis. Nat Genet. 2009;41(2):166–167. [DOI] [PubMed] [Google Scholar]
  11. Gregorutti B, Michel B, Saint-Pierre P.. Correlation and variable importance in random forests. Statist Comput. 2017;27(3):659–678. [Google Scholar]
  12. Hartanto M, Joosen RVL, Snoek BL, Willems LAJ, Sterken MG, de Ridder D, Hilhorst HWM, Ligterink W, Nijveen H.. Network analysis prioritizes DEWAX and ICE1 as the candidate genes for major eQTL hotspots in seed germination of Arabidopsis thaliana. G3 (Bethesda). 2020;10(11):4215–4226. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Ho TK. The random subspace method for constructing decision forests. IEEE Trans Pattern Anal Mach Intell. 1998;20(8):832–844. [Google Scholar]
  14. Ito T, Chiba T, Ozawa R, Yoshida M, Hattori M, Sakaki Y.. A comprehensive two-hybrid analysis to explore the yeast protein interactome. Proc Natl Acad Sci U S A. 2001;98(8):4569–4574. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Jansen R, Nap J.. Genetical genomics: the added value from segregation. Trends Genet. 2001;17(7):388–391. [DOI] [PubMed] [Google Scholar]
  16. Jansen RC, Tesson BM, Fu J, Yang Y, McIntyre LM.. Defining gene and QTL networks. Curr Opin Plant Biol. 2009;12(2):241–246. [DOI] [PubMed] [Google Scholar]
  17. Jimenez-Gomez JM, Wallace AD, Maloof JN.. Network analysis identifies ELF3 as a QTL for the shade avoidance response in Arabidopsis. PLoS Genet. 2010;6(9):e1001100. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Joo Y, Fragoso V, Yon F, Baldwin IT, Kim SG.. Circadian clock component, LHY, tells a plant when to respond photosynthetically to light in nature. J Integr Plant Biol. 2017;59(8):572–587. [DOI] [PubMed] [Google Scholar]
  19. Julca I, Ferrari C, Flores-Tornero M, Proost S, Lindner A-C, Hackenberg D, Steinbachová L, Michaelidis C, Pereira SG, Misra CS, et al. Comparative transcriptomic analysis reveals conserved transcriptional programs underpinning organogenesis and reproduction in land plants. bioRxiv. 2020. [DOI] [PubMed]
  20. Keurentjes JJB, Fu J, Terpstra IR, Garcia JM, van den Ackerveken G, Snoek LB, Peeters AJM, Vreugdenhil D, Koornneef M, Jansen RC, et al. Regulatory network construction in Arabidopsis by using genome-wide gene expression quantitative trait loci. Proc Natl Acad Sci U S A. 2007;104(5):1708–1713. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Kim JY, Song HR, Taylor BL, Carre IA.. Light-regulated translation mediates gated induction of the Arabidopsis clock protein LHY. EMBO J. 2003;22(4):935–944. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Lelli KM, Slattery M, Mann RS.. Disentangling the many layers of eukaryotic transcriptional regulation. Annu Rev Genet. 2012;46:43–68. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Lin F, Fan J, Rhee SY.. QTG-Finder: a machine-learning based algorithm to prioritize causal genes of quantitative trait loci in Arabidopsis and rice. G3 (Bethesda). 2019;9(10):3129–3138. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Lin F, Lazarus EZ, Rhee SY.. QTG-Finder2: a generalized machine-learning algorithm for prioritizing QTL causal genes in plants. G3 (Bethesda). 2020;10(7):2411–2421. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Liu D, Zhao D, Li X, Zeng Y.. AtGLK2, an Arabidopsis GOLDEN2-LIKE transcription factor, positively regulates anthocyanin biosynthesis via AtHY5-mediated light signaling. Plant Growth Regul. 2022;96(1):79–90. [Google Scholar]
  26. Lowry DB, Logan TL, Santuari L, Hardtke CS, Richards JH, DeRose-Wilson LJ, McKay JK, Sen S, Juenger TE.. Expression quantitative trait locus mapping across water availability environments reveals contrasting associations with genomic features in Arabidopsis. Plant Cell. 2013;25(9):3266–3279. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Marbach D, Costello JC, Küffner R, Vega NM, Prill RJ, Camacho DM, Allison KR, Kellis M, Collins JJ, Stolovitzky G; DREAM5 Consortium. Wisdom of crowds for robust gene network inference. Nat Methods. 2012;9(8):796–804. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Mostafavi H, Spence JP, Naqvi S, Pritchard JK. Limited overlap of eQTLs and GWAS hits due to systematic differences in discovery. bioRxiv, 2022.
  29. Nijveen H, Ligterink W, Keurentjes JJB, Loudet O, Long J, Sterken MG, Prins P, Hilhorst HW, de Ridder D, Kammenga JE, et al. AraQTL—workbench and archive for systems genetics in Arabidopsis thaliana. Plant J. 2017;89(6):1225–1235. [DOI] [PubMed] [Google Scholar]
  30. O'Malley RC, Huang S-SC, Song L, Lewsey MG, Bartlett A, Nery JR, Galli M, Gallavotti A, Ecker JR.. Cistrome and epicistrome features shape the regulatory DNA landscape. Cell. 2016;165(5):1280–1292. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Pandey G, Zhang B, Chang AN, Myers CL, Zhu J, Kumar V, Schadt EE.. An integrative multi-network and multi-classifier approach to predict genetic interactions. PLoS Comput Biol. 2010;6(9):e1000928. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Petch J, Di S, Nelson W.. Opening the black box: the promise and limitations of explainable machine learning in cardiology. Can J Cardiol. 2022;38(2):204–213. [DOI] [PubMed] [Google Scholar]
  33. Rockman MV, Kruglyak L.. Genetics of global gene expression. Nat Rev Genet. 2006;7(11):862–872. [DOI] [PubMed] [Google Scholar]
  34. Serin EA, Nijveen H, Hilhorst HW, Ligterink W.. Learning from co-expression networks. Front Plant Sci. 2016;7(444):444. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Skelly DA, Ronald J, Akey JM.. Inherited variation in gene expression. Annu Rev Genomics Hum Genet. 2009;10:313–332. [DOI] [PubMed] [Google Scholar]
  36. Snoek LB, Terpstra IR, Dekter R, Van den Ackerveken G, Peeters AJ.. Genetical genomics reveals large scale genotype-by-environment interactions in Arabidopsis thaliana. Front Genet. 2012;3(317):317. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Szklarczyk D, Gable AL, Lyon D, Junge A, Wyder S, Huerta-Cepas J, Simonovic M, Doncheva NT, Morris JH, Bork P, et al. STRING v11: protein-protein association networks with increased coverage, supporting functional discovery in genome-wide experimental datasets. Nucleic Acids Res. 2019;47(D1):D607–D613. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Terpstra IR, Snoek LB, Keurentjes JJ, Peeters AJ, van den Ackerveken G.. Regulatory network identification by genetical genomics: signaling downstream of the Arabidopsis receptor-like kinase ERECTA. Plant Physiol. 2010;154(3):1067–1078. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Tian F, Yang DC, Meng YQ, Jin J, Gao G.. PlantRegMap: charting functional regulatory maps in plants. Nucleic Acids Res. 2020;48(D1):D1104–D1113. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. van Muijen D, Anithakumari AM, Maliepaard C, Visser RG, van der Linden CG.. Systems genetics reveals key genetic elements of drought induced gene regulation in diploid potato. Plant Cell Environ. 2016;39(9):1895–1908. [DOI] [PubMed] [Google Scholar]
  41. Vogel C, Bashton M, Kerrison ND, Chothia C, Teichmann SA.. Structure, function and evolution of multidomain proteins. Curr Opin Struct Biol. 2004;14(2):208–216. [DOI] [PubMed] [Google Scholar]
  42. Waters MT, Wang P, Korkaric M, Capper RG, Saunders NJ, Langdale JA.. GLK transcription factors coordinate expression of the photosynthetic apparatus in Arabidopsis. Plant Cell. 2009;21(4):1109–1128. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Wentzell AM, Rowe HC, Hansen BG, Ticconi C, Halkier BA, Kliebenstein DJ.. Linking metabolic QTLs with network and cis-eQTLs controlling biosynthetic pathways. PLoS Genet. 2007;3(9):1687–1701. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. West MAL, Kim K, Kliebenstein DJ, van Leeuwen H, Michelmore RW, Doerge RW, St. Clair DA.. Global eQTL mapping reveals the complex genetic architecture of transcript-level variation in Arabidopsis. Genetics. 2007;175(3):1441–1450. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Wong SL, Zhang LV, Tong AHY, Li Z, Goldberg DS, King OD, Lesage G, Vidal M, Andrews B, Bussey H, et al. Combining biological networks to predict genetic interactions. Proc Natl Acad Sci U S A. 2004;101(44):15682–15687. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Yu G. Gene ontology semantic similarity analysis using GOSemSim. Methods Mol Biol. 2020;2117:207–215. [DOI] [PubMed] [Google Scholar]
  47. Zhang H, Zhang F, Yu Y, Feng L, Jia J, Liu B, Li B, Guo H, Zhai J.. A comprehensive online database for exploring approximately 20,000 public Arabidopsis RNA-Seq libraries. Mol Plant. 2020;13(9):1231–1233. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

jkac255_Supplementary_Table1_2_4_5
Click here for additional data file. (7.1KB, zip)
jkac255_Supplementary_Figure_S1
Click here for additional data file. (62.2KB, pdf)
jkac255_Supplementary_Figure_S2
Click here for additional data file. (390KB, jpeg)
jkac255_Supplementary_Figure_S3
Click here for additional data file. (249.7KB, jpeg)
jkac255_Supplementary_Table3
Click here for additional data file. (15.1KB, xlsx)
jkac255_Supplementary_Table6
Click here for additional data file. (16.6KB, xlsx)

Data Availability Statement

The code and data for the analysis and visualization are available at the Wageningen University GitLab repository (https://git.wur.nl/harta003/eqtg-finder). eQTG-Finder prioritization is available at AraQTL (https://www.bioinformatics.nl/AraQTL/; Nijveen et al. 2017).

Supplemental material is available at G3 online.

Articles from G3: Genes|Genomes|Genetics are provided here courtesy of Oxford University Press

RESOURCES