Genome-wide association studies: assessing trait characteristics in model and crop plants

GWAS involves testing genetic variants across the genomes of many individuals of a population to identify genotype–phenotype association. It was initially developed and has proven highly successful in human disease genetics. In plants genome-wide association studies (GWAS) initially focused on single feature polymorphism and recombination and linkage disequilibrium but has now been embraced by a plethora of different disciplines with several thousand studies being published in model and crop species within the last decade or so. Here we will provide a comprehensive review of these studies providing cases studies on biotic resistance, abiotic tolerance, yield associated traits, and metabolic composition. We also detail current strategies of candidate gene validation as well as the functional study of haplotypes. Furthermore, we provide a critical evaluation of the GWAS strategy and its alternatives as well as future perspectives that are emerging with the emergence of pan-genomic datasets. Supplementary Information The online version contains supplementary material available at 10.1007/s00018-021-03868-w.

It was reported on 11 January 2019 that for humans 3730 GWAS studies had been published with a total of 37 730 single nucleotide variations and 52 415 unique SNV-trait associations above a genome-wide significance threshold [1,2]. Analysis of the staggering increase in the number of associations in the time-lapse figure on the GWAS catalog website (https:// www. ebi. ac. uk/ gwas/) suggests that these numbers have likely increased at least threefold demonstrating the tremendous uptake of this method in recent years. Indeed, as evidenced by the numbers given above since the first GWAS for age-related macular degeneration was published in 2005 [3], well over 50 000 associations of genome-wide significance (P < 5 × 10 -8 ) have been reported between genetic variants and common diseases and traits [1]. Among these studies risk loci for a vast number of diseases and traits, including anorexia nervosa [4], body mass index [5], cancers and their sub-types [6,7], coronary diseases [7], inflammatory bowel disease [8], insomnia [9], type 2 diabetes mellitus [10], and schizophrenia [11], have been reported. Indeed, the number of replicable associations is now dramatically higher than those available in the pre-GWAS era [12]. The rapid uptake of GWAS in plants is similar. Indeed, since early studies on flowering time and pathogen resistance [13], single feature polymorphism [14], and recombinant and linkage disequilibrium [15], well over 1000 GWAS studies have now been published in plants [16,17]. The data from many of these have subsequently been uploaded to the AraGWAS catalog database [18]. In this article we will provide a review of these studies in plants splitting them into four major categories: (1) biotic resistance, (2) abiotic tolerance, (3) yield associated traits, and (4) metabolic composition. We will document strategies of validation and cross-validation and outline how results from these studies are being exploited both as a route by which to gain mechanistic understanding of various biological processes and one to improve agriculture. Finally, we outline alternatives to the GWAS approach as well as providing a prospective for its future application. However, before doing so we feel it highly important to provide a brief overview of the technique itself.

The GWAS approach
The aim of GWAS is exceedingly simple-namely to detect association between allele or genotype frequency and trait status. The first step of such analysis is to identify the traits to be scored and select an appropriate study population considering both the size of the population and the amounts of genetic and trait variance that it possesses (Fig. 1). Depending on whether using a novel population or one that is already well studied genotyping may or may not be necessary. It can be carried out using single nucleotide polymorphism (SNP) arrays combined with imputation [19] or via whole-genome sequencing [2]. Association tests are then used to identify genomic regions that associate with the variance of the phenotype of interest at genome-wide significance with meta-analysis often used to increase the statistical power to detect associations. The first GWAS was performed by Klein et al. [3], who identified a variant of the Complement Factor H gene as being strongly associated with age-related macular degeneration. Within the last 15 years it has been powerful in dissecting the genetic basis for variation in a range of complex phenotypes including disease in humans and animals and physiological and agronomic traits in plants [20][21][22][23][24][25][26]. That said population structure and unequal relatedness between individuals can result in spurious associations and thereby false discoveries. To combat this problem considerable effort has been made to statistically account for population structure [27,28]. For example, in mixed linear models (MLM), population stratification is fitted as a fixed effect, while kinship among individuals is incorporated via the

IntegraƟon analysis
Phenotyping Phenotypic traits Yield associated AbioƟc tolerance BioƟc resistance CRISPR/Cas9 VIGs G C Fig. 1 A schematic view of GWAS in plants variance-covariance structure of the random effect for the individual [29,30]. Indeed the MLM method is now firmly established in GWAS since it has proven effective in correcting for the inflation of small genetic effects and controlling bias caused by population structure. Generally such models are carried out with single-locus test, however, multi-locus mixed models have been developed which perform well [31,32]. While also commonly used single nucleotide polymorphism (SNP)-based GWAS suffers from oft-overlooked interactions between SNPs within a gene and also weak signals aggregating within related SNP sets [33]. To limit such problems, haplotype-based GWAS and gene-based GWAS have been developed which has high statistical power to identify causal haplotypes and demonstrated to be able to identify new candidates for complex traits albeit being less capable of detecting QTL than SNP-based GWAS especially so for rare alleles [34][35][36]. All these methods are based on the assumption that phenotype and marked effects follow a normal distribution. Two further developments are worthy of note. The Anderson Darling test is a complementary method, which is particularly useful for moderate effect loci or rare variants and with abnormal phenotype distribution [37] while statistics-based fine-mapping strategies have also been developed [38].
Initial excitement surrounding GWA cooled considerably on the appreciation of the above-mentioned facts that GWAS loci often have small effect sizes and explain only a modest proportion of heritability [39]. However, this missing heritability is, at least as long as large and varied populations as used, in fact rather small. What is clear is that the larger the population and the larger the number of SNPs the greater the chance of a successful result with empirical evidence demonstrating that for each complex trait there is a threshold sample size above which the rate of locus discovery accelerates in GWAS [40,41]. It is important to note, however, that the value of biological insight gained from GWAS is in no way proportional to the strength of association, a fact that provides a strong argument for the value of finding subtle associations in ever larger sample sizes [42]. As stated above genetic variants can be genotyped in many different ways but by far the most predominant are SNP arrays and whole-genome sequencing (see Fig. 1). Given the lowering sequencing costs the latter is beginning to become more frequent. The advantages of SNP arrays, other than their lower costs are the fact that it is highly accurate with a wellestablished pipeline for analysis. By contrast, although less accurate and more expensive whole-genome sequencing provides coverage also of rare variants and even if the sample size is large enough ultra-rare variants. In addition fine mapping is easier with whole-genome sequencing, however, these advantages come at the cost of higher computational costs including a higher multiple testing burden [2]. To offset some of the limitations of SNP-based GWAS sophisticated tools for genotype imputation have been developed which allow genotypes or untyped variants to be predicted. If the size of the reference panel is large enough and a subset is well sequenced this imputation has been demonstrated to be highly reliable [43,44]. Given this fact it is not surprising that both approaches currently retain utility. However, whole-genome sequencing is the gold standard in GWAS [45][46][47] and has the potential to resolve many of the limitations of the method (for example the identification of missed signals, accounting for population stratification, identification of ultra-rare mutations as well as gene-gene and geneenvironment interactions and to explain even more of the missing heritability). We will discuss this in detail when we compare GWAS with other strategies to link genotype with phenotype in Limitations of GWAS an alternative approaches to GWAS below. Having provided a general introduction to the approach above we will, use early case studies in Arabidopsis that span a wide range of phenotypic traits to illustrate it in detail below before providing a more comprehensive overview of its use in other species.

Early studies of GWAS in Arabidopsis
As for many studies in the last 40 years the initial applications of GWAS in plants were in Arabidopsis. The very earliest studies focused on single feature polymorphism [14] and recombination and linkage disequilibrium [15], but a far more diverse range of phenotypes have been studied in the interim. The study of Borevitz et al. used hybrization to a microrarray as a means to assess genomic DAN diversity of 23 ecotypes in comparison to the reference ecotype Col0 allowing assessment of over 77 000 single feature polymorphisms [14]. Similarly, that of Kim et al. analyzed linkage disequilibrium in a sample of 19 Arabidopsis accessions using approximately 350 000 non-singleton SNPs demonstrating the presence of clear recombination hotspots in intergenic regions [15]. Currently, in Arabidopsis results of > 400 GWAS covering an exhausting range of phenotypes are curated in the AraGWAS catalog [18]. To highlight a few recent studies we will focus on growth, metabolism, defense, and evolution of tolerance to abiotic stress [48][49][50][51][52]. Growth and metabolism have been evaluated in association with enzyme activities of primary metabolism [48], while primary [51] and secondary metabolite contents [49,50] have also been studied via the use of metabolomics approaches. All of these studies have provided greater insight into the interplay between metabolism and growth on one hand and defense on the other [53], with both difference in the levels of defense metabolites and altered alleles of ACCEL-ERATED CELL DEATH6 suggesting a trade-off between metabolism and defense. Abiotic stress has also been much studied in Arabidopsis populations with the recent tourde-force work of Exposito-Alonso representing a beautiful example of the power of this approach [52]. These authors evaluated 517 Arabidopsis ecotypes grown in Spain and Germany simulating high and low precipitation at each site quantifying survival and fecundity and thereafter performing a GWAS in the quantified selection coefficients. They observed that a significant proportion of the climate-driven natural selection was predictable form signatures of local adaptation since genetic variants were found in geographical areas with climates more similar to the experimental sites were positively selected. These data thus allowed them to forecast that with the increased frequency of drought and temperature in Europe such positive selection will sweep Northwards across Europe.

Adoption of GWAS in crop species (i) biotic resistance
In the above section we have detailed some studies evaluating biotic stress in Arabidopsis. In crops this is of massive importance with 20-40% yield losses predicted to be caused by biotic interactions annually. While considerable success has been made by breeding efforts-notably the introgression of wild species alleles conferring resistance [73,74]. Critically the collection of broad populations for, among others, the species listed above renders GWAS, an attractive approach for the identification of further genes of interest for this purpose. As can be seen in Supplementary Table 1, there are already a vast number of such studies covering many species. Here, we will highlight only the few summarized in Table 1.
Starting with studies in our major cereals we will describe two studies each for maize and wheat and one for rice before highlighting the possible value of this approach in two less studied crops. The first study in maize used the nested association mapping population to identify 32 QTL with small additive effects on southern leaf blight with many being within or near genes previously shown to be involved in plant disease resistance [56]. More recently, GWAS revealed that the F-Box protein ZmFBL41 which interacts with ZmCAD encoding the terminal enzyme of the monolignol pathway which if active restricts lesion expansion [75]. Similarly, in a GWAS-based study in rice Li et al. found a natural allele of a C 2 -H 2 type transcription factor that confers broad spectrum resistance. Haplotype analysis (which we will return to it below), revealed that this allele exists in 10% of accessions of rice. This allelic variance was associated to an inhibition of H 2 O 2 degradation which the authors postulate is responsible for the observed resistance. In Emmer wheat stripe resistance loci that were associated with field resistance in multiple environments with more than half of these representing novel candidate genes that were not found in linkage mapping studies [76]. Meanwhile, a recent large-scale study in 2 300 bread wheat accessions was used to investigate leaf-, stem-, and stripe-rust diseases with both single-and multi-trait GWAS being applied [77]. Importantly, both studies revealed the utility of small effect QTL in achievement of durable resistance.
Of the less studied species, we would highlight two cassava which is actually the fourth largest crop in terms of production globally [78] and pigeonpea an important smallholder crop in India and Africa [79]. For cassava GWAS for cassava mosaic disease and cassava green mite severity were carried out identifying several novel and previously reported associations. For pigeonpea a pangenome was recently published based on 89 accessions and this will surely be a fantastic resource for future studies. Indeed, since so many natural populations are now established it would seem likely that their use as well as those of biparental and multi-parental populations will likely unlock resistance in a wide range of plant-pest combinations and as such will result in the achievement of durable resistance.

Adoption of GWAS in crop species (ii) abiotic tolerance
Similarly to the above studies aiming to generate more resistant plants considerable research and breeding efforts have been expended on identifying and utilizing allelic variance that confers tolerance to abiotic stresses. As can be seen in Supplementary Table 1, there are already a vast number of such studies covering many species. Here, we will highlight only the few summarized in Table 1 focusing on water and salt stress as well as macronutrient and temperature stress. Arguably, the most important of these is drought stress with yield losses of > 50% being estimated to be due to this stress annually [80]. While water deficiency can devastate crop yields the opposite, i.e., flooding can have the same consequences. The development of varieties of rice that are tolerant of flooding is thus highly desirable.
The identification of haplotypes of the SEMIDWARF1 gene that facilitate this [81] presents an excellent example of the power of haplotype analysis following GWAS studies (an analysis type we will return to it below). Similarly in rice, salt stress has been much researched. Al-Tamanini et al. combined high throughput phenotyping of plant growth and transpiration with high-density genotyping if indica and aus diversity panels containing a total of 553 accessions [82]. This study identified a previously undetected loci for salt stress localizing to chromosome 11, thus, providing new insight into early responses to rice salinity and providing hints as to how breeding could alleviate this problem.
Given that nitrogen fertilizer is often over applied to fields often with catastrophic ecological consequences. There is, thus, a pressing need to develop crops exhibiting high nitrogen use efficiency to reduce fertilizer to move towards a more sustainable agriculture. Tang et al. recently identified the nitrate transporter OsNPF6.1 (HapB) as conferring high nitrogen use efficiency in a GWAS experiment conducted on a rice diversity panel [83] with haplotype analysis identifying that this allele had been lost in over 90% of rice varieties. In a similar vein GWAS was used to investigate phosphate use efficiency in maize [84] with metabolomics being utilized in this study to understand how metabolism is reprogrammed under phosphate limitation. The combined work identified phosphoglucose isomerase activity to be a key determinant of phosphate use efficiency suggesting it to be a strong lead gene for lessening the need of P fertilization [84]. Extreme temperatures also often provoke deleterious effects on crop yield. For this reason, GWAS was recently applied to identify genes underlying cold tolerance in a large 1033 accession rice diversity panel [85]. This study resulted in the identification of five cold tolerance related genetic loci with one loci LOC_Os10g34840 being deemed responsible for cold tolerance at the seedling stage with the cold tolerant allele being present in 80% of temperate japonica accessions but only 3.8% of the indica accessions. By contrast, for high temperature tolerance, GWAS discovered genetic factors associated with four production traits in both heat and drought stress environments in common bean (Phaseolus vulgaris L.) [86].

Adoption of GWAS in crop species (iii) yield associated traits
Having addressed the use of association mapping in resistance and tolerance of plants to biotic and abiotic factors, respectively, above it is important to note that considerable research effort has additionally been placed on elucidating the genetic basis of yield associated traits. As for the above traits we have listed several GWAS studies reporting yield associated traits in Table 1 and provide a more extensive list in Supplementary Table 1. An early study tested almost 5000 lines from the maize NAM population described above to identify numerous small effect QTL with a simple additive model being able to predict flowering time [87]. In addition to flowering time, in rice panicle architecture is a key target of selection. A total of 49 panicle phenotypes were recently assessed in 242 tropical rice accessions allowing the identification of ten GWAS peaks but also demonstrating subtle links between panicle size and yield performance [88]. The complexity of agronomic yield was similarly underlined by a study of 84 agronomic traits in a panel of 809 soybean accessions with many of the loci exhibiting complex pleiotropic effects [89]. In upland cotton a GWAS identified two ethylene pathway related genes as associated with increased lint yield with an analysis of population frequencies revealing that the majority of the elite alleles detected were transferred from a mere three founder landraces [62]. Such analyses are not restricted to cereals with analysis even being carried out in long lived species such as Populus trees [90], as well as sweet potato [91] and GWAS confirming the Lin5 association with agronomic yield in tomato [25] that had previously been identified by linkage mapping [92]. It is perhaps not unexpected that the QTL for yield associated traits seem generally not to be conserved across species.

Adoption of GWAS in crop species (iv) metabolic composition
Combining the developments in sequencing with those in mass-spectrometry-based analytical systems, has rendered understanding of the genetic architecture of metabolism far easier than it was previously [33,[93][94][95]. Indeed the immense metabolic diversity of plants has made the ideal models for dissecting the genetic bases underlying the regulation of the metabolome with studies progressing from analysis of mutant libraries [96,97], and the analysis of gene families [98,99] via the comparison of sister species [100] and species series within taxa [101] to linkage mapping, and association mapping based on next-generation sequencing have been applied to metabolomics studies [33]. By contrast to the QTL for agricultural performance described above, genetic variants controlling natural variation in metabolite accumulation are easier to identify due to both the tremendous diversity apparent across experimental populations [20,[102][103][104][105] and the high accuracy of evaluation of metabolite content [95]. As mentioned above a wide range of examples are now published both in cereal and non-cereal crops ( Table 1 and Supplementary Table 1). Due to space limitations we limit our discussion to ten of these examples. In maize, GWAS was used to quantify metabolite contents of nearly 1000 mass features in over 700 lines and further allowed the association of metabolite features with kernal size [55] while a more recent study identified four times as many features paying particular attention to the benzoxazinoids and hydroxycitric acids [103]. Earlier a groundbreaking highly comprehensive study on maize kernel oil identified 74 associated loci of which 26 were found that could explain up to 83% of the phenotypic variation using a simple additive mode.
Maize kernel oil is a valuable source of nutrition. In a seminal study, Li et al. examined the genetic architecture of oil accumulation in maize by GWAS using 368 maize inbred lines characterized to contain in excess of 1 million SNPs. In the process, they identified 74 loci associated with kernel oil levels and fatty acid composition. They validated more than half of these in a linkage mapping population and 26 of the conserved loci were annotated as enzymes of oil biosynthesis and could explain up to 83% of the phenotypic variation in this trait [106]. Similarly in rice, secondary metabolism data of 175 accessions identified 323 associations among 1 3 143 SNPs and 89 metabolites. While a comparative analysis between maize and rice demonstrated a considerable amount of shared loci associated with metabolites common to both species [20], but of course could not provide information with regard to species-specific metabolites or for that matter genes [33]. The use of this approach in wheat and barley has allowed the definition of the flavonoid biosynthesis pathway in the former and a novel metabolite, thereof, that confers UV-tolerance in hulless barley, respectively. In tomato, GWAS was used in concert with metabolite profiling and taste panels to characterize the genetic architecture of tomato fruit taste [25] and with metabolic and transcript profiling to characterize the changes in the metabolome that occurred during the domestication and improvement processes [26] while a combination of GWAS, a multi-parental breeding population and transgenic lines was used to characterize the control of vitamin E levels in this fruit [107]. To summarize, metabolic GWAS has proven highly informative not only as a means of identify lead genes for engineering of specific metabolite contents but also in beginning to define the biological function of specific metabolites [95]. However, in certain species such as citrus the use of GWAS is not yet tractable most likely due to population structure issues (unpublished), and this fact is important to keep in mind before carrying out labor-intensive studies, on a new species-irrespective of the phenotype studied.

Validation of candidate gene function
Despite the strong theoretical foundation we discuss above and considerable efforts being taken to address population structure and employ strict probability cut-offs, false-positive associations will still occur due to the enormous number of statistical inferences and other factors which are not taken into account by the simplicity of the approach [17,108,109]. As a consequence independent biological validation is required, however, often not provided [17]. That said two forms of validation have been employed in several instances (i) the validation of associations in independent populations or (ii) validation by targeted viral-induced gene silencing, transgenesis and gene editing experimentation. Cross-population validation is currently largely achieved by integrating association mapping in diverse panels or linkage mapping in RIL population(s) or F2 populations. For example, in the recent cloning of ZmCCT9, a QTL which affects maize flowering time [110], the locus was simultaneously identified by NAM [87] and maize-teosinte RIL populations under association and linkage mapping. Moreover, the causal allele-an InDel of a harbinger-like transposon-has also been identified in a 513 line association panel [111] a fact that was cross-validated in the two populations used to map the locus. In a similar example, rice chlorophyll content was mapped in a panel of 529 individuals followed by three customized F2 populations [112]. Other such examples are the metabolomes of maize [113] and in independent studies the QTL underlying total soluble solid content [92,113] and alterations in the metabolome [26,93] in tomato and the exquisitely controlled study mentioned above which used GWAS, multi-parental breeding populations and transgenics to confirm QTL for tocopherol contents [107]. The increasing availability of populations which have been characterized should massively increase or capacities to do such experiments which will undoubtedly massively boost our confidence in the results of association mapping studies. In this vein, it is important also to note also the value of cross-species analysis which has already been implemented in cereals [20,114,115] and would probably prove tractable in other agronomically important families such as the Brasicacae, Solanaceae, and legumes. Rather than employing the cross-validation approach which can prove incredibly time and labor intensive several other more direct approaches have been taken. For example, the confirmation of many metabolic QTL has been provided by the reduction of the expression of candidate genes via virus-induced gene silencing [93,95,116] or alternatively via their transient or inducible expression [20]. Given that the repertoire of species amenable to both methods are currently being considerably expanded. While these are great for select candidates the promise of clustered regularly interspaced short palindromic repeats (CRISPR)/CRISPR associated protein 9 (Cas9) mutant libraries such as those set up for rice [117,118] and more recently maize [119] should greatly accelerate the functional confirmation of causality. Like the VIGS and transient expression methods, the range of plant species for which multiple publications on the use of CRISPR has seen a steep increase in recent years [119,120].

Limitations of GWAS an alternative approaches to GWAS
Despite the great success of the method as evidenced by the wealth of information described above (and in the Supplementary Table 1), GWAS currently has clear limitations the major of which being issues concerning population structure and low-frequency causal alleles leading to false negative results [121]. For example, given that flowering time is a typical adaptive trait and is always confounded (i.e., highly correlated) with population structure, only one gene (ZmCCT) was revealed for flowering time using a diverse association mapping panel consisting of 500 inbred lines [122]. It is widely accepted that many false negatives occur for such confounded traits when correcting for population structure in GWAS [17,123]. Another example is the demonstration that only five inbred lines in a population of 1 3 527 (< 1%) possess functionally alternative alleles at the Brachytic2 locus for plant height [124] rendering it impossible to identify this locus using routine association mapping analysis. Similarly in rice, causal alleles within most of the cloned yield related quantitative trait loci (QTLs) are at low frequency in diverse germplasms (1% for GS3, [125]; 2% for Ghd7, [126][127][128]; 2% for qGL3, [129]; 6% for TGW6, [130]). Two routes to tackle these issues have been suggested either the development of novel statistical methods for the exploration of rare functional alleles [131][132][133] or alternatively employing artificially designed populations to balance allelic frequencies and thereby control population structure [87,[134][135][136]. Given that these have been reviewed in depth recently [17,[137][138][139]we will not discuss them in detail here.
In addition to the above issues, sometimes non-causative loci show more significant associations in GWAS than the causative ones meaning the causative genes may be distant from the GWAS peaks. Such an occurrence has been reported in a number of plant studies including studies in Arabidopsis [140,141], sorghum [142], and tomato [143]. Such misleading associations are sometimes known as synthetic associations and are presumed to be caused by linkage drag caused by linkage disequilibrium between common tagged markers and rare causative variants [17,144]. This may in turn explain the so-called missing heritability issue of GWAS. That said some causes do not follow the rare-allele assumption but trait variation rather appears to be caused by multiple alleles within one gene [34,142]. Given that mutation constantly generates new variants, multiple independent alleles within one gene leading to the same phenotype could be common. As we state above haplotype-or gene-based methodologies, therefore, have high potential for identifying such situations. That said current haplotypebased association mapping remains imperfect [145] and, moreover, is particularly challenging in plants [17]. Thus improving haplotype analyses will likely prove highly beneficial both at the understanding of the underlying genetics as well as its functional physiological consequence.

Current and future perspectives for GWAS
The power of genome-wide association studies have successfully identified enormous number of loci associated with phenotypic, expression, and metabolic traits in multiple species. Although, the genetic factors underling some of these associations have been characterized. The vast majority are remain unexplained. The development of next-generation sequencing and bioinformatics tools greatly improved and currently implemented to decipher the genetic diversity of targeted traits. This recently supported by multi-omics data analysis to enhancing our understanding of phenotypic diversity and its corresponding genetic basis. Combined analyses of phenotypic and transcriptomic data have been utilized to dissect the genetic basses of various metabolic and phenotypic traits see [146]. Moreover, the developments of molecular biology techniques (e.g., CRISPR/Cas9, over-expression, or genetic complementation) have greatly accelerated the biological functions of the causative genes behind the GWAS hits. Currently, the cross-validation by combing association and linkage (F2, RILs) mapping has already been implemented in crop [25,147]. Finally, despite molecular and genetic validations are the reliable ways to validate the GWAS results, there are still accompanying challenges need to take into consideration, such as; epistasis, heterosis and environmental factors. Once such factors are assembled, it will improve our chance of understanding the genetic regulation of complex traits, and provide viable targets for crop improvement and breeding.

Declarations
Conflict of interest The authors declare that they have no conflict of interest, Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.