2.1 Natural Variation of Rosette Leaf Epidermis Traits

We analyzed 310 natural A. thaliana accessions, mostly with available detailed location data and climate parameters of their sites of origin (Supporting Information S2: Table 1) to gain insight into the genetic basis of variability in trichome shape, size and density, epidermal autofluorescence and epidermal metal deposition patterns. We recorded 7 continuous quantitative and 11 categorical semiquantitative (ordinate) or qualitative parameters (Table 1; Bezvoda 2023) that exhibited readily noticeable variabiLity among the accession and were amenable to semi-high-throughput analysis with good interobserver replicability.

The first group of parameters with continuous distribution (Figure 1) describes quantitative aspects of trichome shape, that is, size-related parameters (area, perimeter, overall length and stem length) and descriptors of shape complexity (circularity and solidity). Other morphological features, namely trichome density (Figure 2A,B) and branch number (Figure 2C), were evaluated semiquantitatively.

The next group of parameters reflects structural and chemical properties of the trichome cell walls, affecting trichome autofluorescence upon damage, estimated as a semiquantitative categorical variable (Figure 2D,E) and callose content, evaluated semiquantitatively on a continuous scale in a manner capturing also the spatial distribution of callose (Figure 2F,G). Obvious variability in the background epidermal autofluorescence outside trichomes was recorded as a categorical semiquantitative variable (Figure 2D). We introduced another categorical qualitative variable to describe visible differences in the colour of trichome and background fluorescence (Figure 3A).

A third group of categorical semiquantitative parameters captures the spatial distribution of metal ion deposits, as detected by dithizone staining, in trichomes and other epidermal cell types (Figure 3B; for description of parameters see Table 1).

Since we noticed that trichomes of some accessions resist mechanical detachment during staining for callose, we recorded this feature as a qualitative parameter ‘shaveproof’.

All seven continuous variable parameters exhibit smooth distribution with a single roughly symmetrical maximum (Figures 1B,D,F, 2G, 4A), except callose content, where most accessions have very Little if any callose in trichomes, resulting in an asymmetric distribution. A roughly symmetrical single maximum distribution, in some cases noisy, was observed also for the semiquantitative categorical variables of trichome autofluorescence, epidermis autofluorescence, trichome density and trichome branch number. The ‘noisy single maximum distribution’ may be explainable by under-representation of visually attributed intermediate values, perhaps reflecting observer uncertainty rather than genuine differences. For several other parameters (metal gradient, metal surround, metal stomata) the distribution was asymmetric with most accessions exhibiting low or no signal. Remaining parameters either varied between only two values (autofluorescence colour, shaveproof) or showed a possibly bimodal distribution, such as in the case of metals at trichome base and metal rings (Supporting Information S1: Figure S1).

Not surprisingly, some quantitative descriptors of trichome shape were mutually correlated (Figures 4A and Supporting Information S1: Figure 1). Trichome area, perimeter and overall length, which can all be viewed as measures of trichome size, exhibited strong or very strong positive correlation. Weak to moderate positive correlation with these parameters was found also for trichome stem length. Circularity (proportional to the ratio of an object's area and squared perimeter) and soLidity (ratio of the area of the examined object and its convex hull), which both achieve higher values for less intricate shapes, were mutually moderately positively correlated, and negatively correlated with the size-related parameters. While there were some cases of non-negligible correlation among the categorical variables, or between categorical and continuous ones, such correlations were generally weak except of moderate positive correlation between the frequency of metal rings and intensity of metal gradients at the trichome stems, as well as moderate positive correlation between strength of mechanical attachment (the ‘shaveproof’ trait) and trichome callose content, which was also moderately negatively correlated with trichome length. Remarkably, all the strong or moderate correlations were statistically significant (Supporting Information S1: Figure 1).

For a subset of traits including representatives of the mutually correlated trait groups, heritability was estimated from data collected separately for individual plants (Supporting Information S2: Table 2). The broad sense heritability values, ranging between 0.58 and 0.94, indicate a strong genetic contribution to the observed variability (Supporting Information S2: Table 3).

Principal component analysis (PCA) of the 18 recorded parameters did not reveal any obvious clustering of accessions when considering the first three principal components (PCs), which, however, cumulatively explained only less than 43% of total variability (Figure 4B). Analysis of PC loadings indicated that PC1 is positively correlated with the mutually correlated parameters reflecting trichome size, and negatively with circularity and solidity, PC2 correlates either positively or negatively with specific patterns of metal staining, and PC3 mostly reflects callose content and strength of mechanical attachment—that is, two parameters related to cell wall structure (Figure 4C). The PCA results thus support the overall pattern of continuous variability of traits addressed in our study, and also suggest that trichome shape, metal staining and cell wall organization might be determined by distinct sets of genes.

We next examined possible relationships between the 18 phenotypic traits determined in our study and 35 standardized climate variables of the sites of origin of our Arabidopsis accessions, obtained from the CliMond database (Kriticos et al. 2012). Out of the 630 possible trait and cLimatic variable combinations, only 81 exhibited statistically significant non-negligible positive or negative correlation, typically in the weak range (Supporting Information S2: Table 4, Figure 5A). The only exception was a negative correlation between trichome stem length and the maximum temperature of the warmest week (the Bio05 variable), which fell into the moderate range (sensu Schober, Boer, and Schwarte 2018; Figure 5B). The biological interpretation of this observation is not immediately obvious.

2.2 GWAS Analysis Identifies Loci Associated With Some Epidermal Traits

The phenotypic data were subjected to association analysis using the publicly available GWAPP GWAS platform (Seren et al. 2012) as described in Materials and Methods, aiming first towards identification of all single nucleotide polymorphisms (SNPs) within the transcribed section of genomic DNA, including the CDS, introns and the 5ʹ and 3ʹ UTRs. The numbers of identified significantly associated loci varied among traits on the scale from zero to several hundred, with trichome stem length and guard cell metal accumulation yielding the most candidates (Tables 2 and Supporting Information S2: Table 5). No significant SNPs were found for trichome area, metal ring, metal staining of trichome bases, and resistance towards mechanical detachment, and only one or a few presumably silent polymorphisms (i.e., synonymous mutations or SNPs located in an intron or in untranslated mRNA regions) were identified for overall trichome length, number of branches and the presence of a metal gradient. These traits were therefore not included in subsequent analyses.

The distribution of SNP types for individual traits with multiple SNPs exhibits some differences, although they are not major (Supporting Information S1: Figure 2). In particular, SNPs altering CDS length were found only for traits with overall large SNP numbers, reflecting low probability of polymorphisms affecting start and stop codons. Subsequently, we focused only on loci with SNPs changing the predicted protein product sequence, in most cases due to a missense mutation, and treated all such SNPs as a single category. The full list of genes affected by such SNPs for each trait is provided in Supporting Information S2: Table 6. In total, 1547 different loci exhibited association with at least one of the studied traits, corresponding to ~5.6% of all Arabidopsis genes.

For further analyses, we divided the traits with significantly associated SNPs into four groups: those affecting trichome shape (circularity, soLidity, perimeter and stem length), trichome distribution on the surface of the leaf (i.e., trichome density), wall composition-related parameters (epidermis and trichome autofluorescence, autofluorescence colour, callose content) and metal accumulation (metal staining of stomata and trichome surroundings). While some loci associated with several of the mutually correlated cell shape traits, there was only a very Limited, if any, overlap between candidate lists for traits from the wall composition-related and metal accumulation-related groups. However, polymorphisms in some loci were identified as associating with traits from more than one category (Figure 6A, Supporting Information S2: Table 7).

Assuming that SNPs in transposon-borne genes are unLikely to cause phenotypic differences, we examined what fraction of our candidates corresponds to CDSs from mobile genetic elements to gain insights into the extent of non-specific background. The fraction of transposon-derived loci was significantly depleted compared to whole genome values for the sum of all GWAS candidates, as well as for those associated with shape-related traits or trichome density. However, while depletion of transposon genes was also noticeable for wall-related traits, it was only marginally significant, possibly due to a relatively low number of loci, and practically no depletion was observed for metal-related traits (Table 3). Sets of genes associated with multiple loci usually contained fewer transposon-derived loci than those found on the basis of a single trait (Figure 6B).

2.3 Multiple GWAS Candidates Encode Proteins With Known or Suspected Epidermal Roles

Next, we examined the traits related to trichome development (i.e., the shape traits, density and the cell wall parameters that largely reflect trichome properties) for enrichment of genes abundantly expressed in trichomes either at the mRNA (Jakoby et al. 2008) or protein (Huebbers et al. 2022) levels. While we did not detect significant enrichment or depletion of these genes in our candidates lists, we identified multiple genes associated with trichome related traits that were abundantly expressed in trichomes, consistent with their participation in trichome development. In some cases, available annotations suggest function in membrane trafficking, biogenesis of cell wall components or turgor generation and maintenance. This might point to possible mechanisms whereby variation in the responsible loci may affect cell growth or cell wall development and generate the observed phenotypic diversity, either by direct modification of trichome morphogenesis or by modulating epidermal cell expansion and hence trichome density (Table 4). Prompted by the finding of several genes engaged in cuticle or epidermal wax biosynthesis, that is, traits often linked to trichome development (see Berhin, Nawrath, and Hachez 2022) among these loci, we examined the rest of our candidates for overlap with a list of 87 putative cuticle and wax biosynthesis genes identified on the basis of their known activities and/or expression patterns (Suh et al. 2005). However, we did not find any additional loci in this manner beyond two genes that were also highly expressed in trichomes—namely AT1G01600/CYP86A4 and AT1G76690/OPR2 (see Table 4).

Because the loci associated with metal accumulation-related traits mainly reflect stomatal metal deposition, we examined the metal-related candidate list for overlap with a published set of 63 genes upregulated in guard cells (Leonhardt et al. 2004), as well as a list of 1508 proteins repeatedly identified in guard cell proteome (Zhao et al. 2008). We found no overlap between our candidate list and the list of genes highly transcribed in guard cells, but identified 21 loci from the proteome set, all of them associated with metal accumulation in the guard cells and some encoding proteins engaged in ion or other solute transport, membrane trafficking, cell wall biogenesis or gene expression, that is, processes possibly relevant for the observed trait (Table 5). However, there was no significant enrichment of guard cell-expressed genes among our candidates, similar to the trichome case.

To further verify the performance of our GWAS screen, we inspected the candidate gene lists for genes previously reported to affect the studied traits. While such a search cannot be exhaustive, we found at least five previously characterized relevant genes (Table 6), including loci coding for two cytoskeletal proteins CLASP and SCAR/DIS3, as well as the EXO84B gene, encoding a subunit of the exocyst complex that regulates targeting of membrane vesicles. All three genes exhibit mutant phenotypes involving trichome shape alterations, and all associated with trichome shape traits. In addition, the gene for another exocyst subunit, EXO70H4, documented to participate in the deposition of metals in the trichome cell wall, was picked up as associating with guard cell metal deposits. The fifth gene, GFS9/TT9, is known to affect cell wall autofluorescence and was found in association with trichome density and autofluorescence traits. Notably, five SNPs affecting this gene often occurred together, constituting a quintuple substitution minor allele associated with reduced trichome density and decreased autofluorescence.

Additional 80 genes with previously reported roles in cytoskeletal organization, membrane trafficking, cell wall biogenesis, tissue patterning or other relevant functions were also found among the candidates, including several genes affecting aspects of epidermal development unrelated to traits addressed in our study (Supporting Information S2: Table 8).

These findings suggest that, despite the lower efficiency of the metal trait-associated gene search, as evident from the lack of transposon gene depletion, the set of candidates found in our screen contains at least a sizeable proportion of genes biologically relevant in the context of leaf epidermis differentiation.

2.4 Some GWAS Candidates Exhibit Environmentally Correlated Allele Distribution

To obtain clues to the mechanisms underlying the observed correlation between trichome stem length and the maximum warmest week temperature of the accession's location of origin (the Bio05 variable), we further examined the set of genes exhibiting SNPs associated with trichome stem length. Since the full List of over 600 stem length-associated candidates was too large for a detailed study, we focused on the 18 genes that were at the same time abundantly expressed in trichomes or encoded proteins with known roles in trichome development (see Tables 4 and 6).

For twelve of these genes, at least 30 studied genotypes (i.e., a tenth of our accessions set) harboured the minor allele. The distribution of Bio05 values for these loci in all but one case exhibited a statistically significant difference between the major and minor alleles, indicating an environmentally correlated bias in allele distribution (Figure 7A), which might suggest natural selection against the minor alleles in hot habitats. Among the affected genes were loci encoding the microtubule-associated protein CLASP, known to have pleiotropic roles in cell expansion and cell division (Ambrose et al. 2007; Pietra et al. 2013), and SYT1, coding for a synaptotagmin engaged in processes related to stress tolerance and immunity towards pathogens (Schapire et al. 2008; Yamazaki et al. 2008; Levy, Zheng, and Lazarowitz 2015; Kim et al. 2016). It this thus conceivable that the selection was operating on some function(s) of these genes unrelated to epidermal morphogenesis.

For all genes with significant environmentally correlated allele distribution, populations harbouring minor alleles originated from multiple sites, i.e. there was no evidence of spatial restriction due to a founder effect. Even variants only found in Scandinavian accessions occured at two or more geographically distant sites (Supporting Information S1: Figure 3). This further supports the relevance of environmental factors and possible natural selection.

Since the GFS9/TT9 gene displayed an unusual co-occurrence of five SNPs, as described above, we also examined the geographic substitution of the Quintuple substitution minor allele. Because the five contributing substitutions were found in mere six accessions of our experimentally characterized collection, we analyzed the whole set of genotypes from the Ensembl database (Yates et al. 2022). Populations carrying the quintuple SNP allele often originated from mountainous regions, and the average altitude of their sites of origin was significantly higher than for the reference variant (Figure 7B). Correspondingly, a significant difference was found in multiple climatic variables from the standard BioClim set, especially in parameters reflecting typical features of high altitude biotops, such as increased temperature seasonality, greater difference among temperature and precipitation extremes, and increased solar radiation (Supporting Information S1: Figure S1). The presence of the quintuple substitution allele across three continents is consistent with its selective advantage under high altitude conditions.

2.5 Overrepresentation of Some Gene Families Among GWAS Candidates Suggests New Gene Functions

We next examined the sets of loci associated with the four trait groups (trichome shape, trichome density, wall-related traits and metal-related traits) for Gene Ontology term association. No enrichment or depletion of any GO categories was found on the p = 0.01 significance level, while on the p = 0.05 level, a 2.73x enrichment of the Transmembrane signal receptor protein class was found among shape-associated loci, and a 12.22x enrichment of the snoRNA-binding molecular function category was observed among metal trait-associated loci. The biological significance of the latter, however, remains unclear in Light of the apparent low reliability of metal-related candidate detection (see above).

While inspecting our candidate gene Lists, we noticed the presence of multiple members of some gene families, including those encoding evolutionarily conserved domains of unknown function (DUF). We thus examined several such multigene families for possible enrichment among candidates for individual trait groups. Our enrichment analyses included all seven DUF families represented by two or more genes in at least one trait group, as well as five large gene families containing multiple candidates associated with at least one trait group, namely the F-box proteins (568 paralogs in Arabidopsis), receptor-Like protein kinases (307 paralogs), cytochrome P450 isoforms (256 paralogs), cysteine/histidine-rich C1 domain proteins (153 paralogs) and formins (FH2 proteins, 21 paralogs).

For 9 out of these 12 gene families, we found in at least one category of GWAS candidates a reliable, significant at least 2× enrichment compared to whole genome abundance (Figure 8; Supporting Information S2: Table 9). These include receptor-Like kinases (minor but significant enrichment in shape traits association), cytochrome P450 (enrichment in cell wall fluorescence-related traits, compare also Table 4), cysteine/histidine-rich C1 domain proteins (enrichment in shape and metal accumulation-related traits), formins, DUF674, DUF784 and DUF1985 proteins (enrichment in metal accumulation traits), as well as the DUF1261 and DUF3741 families that were found in association with trichome shape traits.

Making use of available previously characterized loss of function mutants affecting all three formin genes associating with the stomatal metal deposition, that is, FH1, FH13 and FH14, we performed single-blinded evaluation of epidermal metal deposition in these mutants and corresponding wild type plants. The results suggest a tendency towards increased stomatal metal accumulation especially in the fh1-1 T-DNA insertion mutant of the formin gene FH1, and possibly also in the CRISPR-generated loss of function allele of the same gene (Figure 9). Although none of the observed between-genotype quantitative differences was statistically significant, this observation is consistent with a possible contribution of formins to epidermal metal deposition that would deserve further attention. It also suggests that also the other observed gene family enrichments may be relevant.

2.6 Associations Among Candidate Genes Indicate Possible Functional Modules

To uncover functional relationships among genes associating with the studied traits, we scanned the candidate Lists for mutual links recorded in the STRING database (Szklarczyk et al. 2023) that registers both experimentally documented and predicted interactions among genes, such as transcriptional co-regulation, clustering on the chromosomes, epistasis, or proteins, such as mutual binding.

Groups of associated genes were found for all four trait groups (Figures 10 and Supporting Information S1: Figure 5, Supporting Information S2: Table 10), but only for trichome density and metal accumulation the numbers of interactions were significantly greater than predicted for a random gene selection (Supporting Information S2: Table 10). Inspection of the found clusters, however, reveals biologically meaningful associations also among candidates Linked to the remaining trait groups. For example, among the shape-associated genes, cluster 15 consists of Exocyst complex subunits. For density and metal deposition, the clusters notably included mainly genes participating in nuclear functions such as chromatin organization, gene expression, RNA processing or nuclear transport.

Since we noticed that some genes are shared by clusters found for two or more trait groups, we next performed an analogous search for inter-gene associations among candidates associating with multiple phenotypic trait groups. With somewhat less stringent criteria than those used in the initial interaction searches, this group of genes also exhibited significantly more mutual associations than a random gene selection (Figure 11; Supporting Information S2: Table 10), with one large and two smaller clusters comprising mostly genes associated with shape traits, as well as a small cluster consisting of genes linked to trichome density. We believe that these findings may provide a starting point for exploration of novel regulatory pathways engaged in Arabidopsis epidermal differentiation.

4.1 Plant Material

A set of 310 natural Arabidopsis thaliana accessions from the 1001 Genomes collection (Alonso-Blanco et al. 2016), has been included in our study (Supporting Information S2: Table 1). Predominantly, genotypes of North-European origin were chosen to reduce the effect of geographical differences in population structure and related parameters reducing GWAS efficacy (see Gloss et al. 2022). Climate variables data for the locations of origin vere retrieved in the form of standardized BioClim variables from the CliMond database (Kriticos et al. 2012).

Seeds were steriLized with chlorine gas, stratified in tubes at 4°C for 3 days in the dark and sown into standard soil substrate pre-treated with Confidor to prevent blackfly infestation. Plants were grown in growth chambers under controlled long-day conditions (16 h light, 23°C/8 h darkness, 18°C; relative air humidity 60%). Three seedlings per genotype were planted into individual pots at the first true leaf stage. Leaves were collected at the age of 5 weeks (at which point plants of some genotypes were beginning to bolt).

For mutant genotype verification for selected formin-encoding genes, homozygous plant lines carrying previously characterized loss of function mutants fh1-1 (SALK-032981; Rosero, Žárský, and Cvrčková 2013), fh13-1 (SALK_064291C; Kollárová, Baquero Forero, and Cvrčková 2021), and fh14-1 (SALK_058886; Li et al. 2010), available from NASC (RRID: SCR_004576), as well as fh1:CRISPR (generated in our laboratory (see Cifrová et al. 2020), with corresponding congenic wild type plants were used.

4.2 Leaf Sample Processing

Three approximately fingernail-sized mature, non-senescent rosette leaves per plant (i.e., nine leaves per genotype) were harvested, resulting in three pooled samples, each containing one leaf from each individual. Samples were stored for at least 2 days before further processing or observation.

For trichome shape evaluation and for determination of callose content in trichomes, leaves were harvested into a tube containing 1x phosphate-buffered saline (PBS) and 100 mM EGTA. Trichomes were isolated and stained for callose by aniline blue as described previously (Kulich et al. 2015). In genotypes where trichomes failed to detach from the leaves (further referred to as ‘shaveproof’), whole leaves were stained for callose using an analogous procedure (Kulich et al. 2018). This in situ staining was also used for additional documentation, such as the photos shown in Figure 1.

For visualization of autofluorescence and for determination of additional morphological parameters, leaves were pressed adaxial side up between two microscope slides and left to dry up to induce breakage of trichomes resulting in increase of autofluorescence (Kulich et al. 2018).

For visualization of metal content, leaves were collected into a tube containing 3 mL of acetone and stored for at least several days. On the day of observation, they were stained by addition of diphenylthiocarbazone (dithizone) reagent freshly prepared from 1.5 mg of dithizone, 1 mL of distilled water and one to two drops of glacial acetic acid (Seregin and Ivanov 1997). After at least 1 h but not more than 6 h of staining, leaves were briefly washed in distilled water and mounted in water on microscopy slides for observation.

4.3 Microscopy and Imaging

Microscopic images were acquired using a Nikon Eclipse 90i fluorescence microscope equipped with PlanApo 4x/0.2 objective and Nikon DsFi 2 camera as described previously (Kulich et al. 2018). For autofluorescence and callose fluorescence detection, UV-B excitation was employed; in case of callose staining, an additional image was acquired in polarized light for trichome visualization. Leaves stained for metal detection were imaged using bright field settings.

If whole leaves were photographed, tilling images covering at least a half of the leaf lengthwise were automatically stitched from multiple frames, which were generated as maximum projections of three images with 50 μm of Z distance, using the Nikon Imaging Software (NIS Elements AR). Additional image processing was performed using the Fiji platform (Schindelin et al. 2012).

4.4 Epidermal Phenotypes Determination

An overview of phenotypic trauts analyzed in our screen is provided in Table 1.

The first group of parameters was determined from photos of callose stained trichomes. At least 25 interactively selected clearly visible trichomes from three leaves of three plants (detached, or, in shaveproof genotypes, 8–9 in situ trichomes per leaf, selected across its diagonal) were evaluated. Trichome shape parameters (area, circularity, soLidity, length and perimeter) were determined from binary images generated by thresholding polarized light photos using the Yen method as implemented in the Fiji software; the procedure was partially automated using recorded macros to increase throughput, with individual trichome selection being the only manual step. Trichome length was measured manually as the longest trichome dimension. Average parameter values from at least 25 trichomes were recorded for these parameters. Callose content was measured as the fraction of trichome area stained for callose, determined using built-in functions of Fiji from binary images generated by thresholding fluorescence and polarized light photos using the Yen method, again with the aid of recorded macros and manual selection of individual trichomes. The reported value for each accession is the sum of percentage values from 25 trichomes.

The second group of parameters was determined from fluorescence microscopy images of dried whole leaves, taken in the autofluorescence channel. Trichome stem length was measured using the NIS elements software with manual endpoint identification; average values from at least 25 clearly visible trichomes from 3 leaves, selected 7–9 per leaf across its diagonal, are reported. Additional parameters (trichome autofluorescence, overall leaf epidermis autofluorescence, autofluorescence colour, trichome density and typical number of trichome branches) were visually categorized (see Table 1).

A third group of parameters reflects visually detectable patterns of metal accumulation, categorized as described in Table 1. This group includes the presence of a metal-enriched Ortmannian ring at the bottom part of the trichome stem (Kulich et al. 2015), presence of a visible metal staining gradient in the trichome stem, presence of metal staining at the trichome base, diffuse epidermal staining outside trichomes and staining of stomatal guard cells.

The final categorical parameter, denoted as ‘shaveproof’, reflects the resistance of trichomes against mechanical detachment during staining for callose.

All categorical values were based on the typical appearance of the epidermis of three leaves. Glabrous leaves were attributed the value ‘0’ for all categorical variables. In theexceptional cases where noticeable differences were seen among the three leaves of a given accession, the majority phenotype was recorded.

Phenotype data for a subset of accessions and traits have been independently rescreened by different team members for the purpose of broad sense heritability estimation (Supporting Information S2: Table 2); this also documented good inter-observer replicability of phenotype determination (Supporting Information S1: Figure 6).

4.5 Phenotypic Data Processing and Initial Analyses

Primary phenotypic data were assembled into a structured spreadsheet and deposited in the pubLic AraPheno database (Togninalli et al. 2020) as Study No. 126 (Bezvoda 2023). Broad sense heritabiLity (H2 = VG/VP), that is, the proportion of phenotypic variation (VP) due to genetic variation (VG) estimated from the between- and within-Line phenotypic variance, was calculated from trait values of all individuals in this data set (Alonso-Díaz et al. 2021).

Statistical analysis of phenotypic data and phenotype/environmental parameter correlation analyses were carried out by the Pandas data analysis toolbox (Reback et al. 2021) and pair plots of scatterplots were created by the Seaborn visualization tool (Waskom 2021) v. 0.11.2. Correlation strength is reported using criteria from Schober, Boer and Schwarte (2018).

Principal component analysis (PCA) has been performed using the PAST software (Hammer, Harper, and Ryan 2001) v. 4.11 using the correlation matrix method. All quantitative or semi-quantitative parameters were treated as ordinal variables, the qualitative ‘autofluorescence color’ and ‘shaveproof’ traits were handled as nominal.

4.6 GWAS Analyses and Genomic Data Processing

Phenotype values for each trait were uploaded to the online GWAS application (Seren et al. 2012; Seren 2015) and correlation analyses were run at 5% FDR threshold with all possible combinations of input data transformations and statistical methods available with default settings except the minor allele count (MAC) value that has been lowered to 5. The 1001 Full sequence Data set (TAIR v. 9) was used as the source of genotype information. After initial manual exploration of resulting Manhattan plots, the process of analyzing data was automated.

Raw result data files were downloaded manually and used for subsequent steps. All SNPs identified as significant by at least one method were considered. Additional information about each significant hit was acquired from the GWAS application by web scraping using the Python programming language v. 3.8 (Python Software Foundation 2024) and driver for Google Chrome web browser (Google 2024). Locus annotation was based on TAIR9.0 to maintain compatibility with the 1001 genomes data set. All intergenic hits were discarded from subsequent steps of data processing. Hits were categorized according to the position of the SNP in corresponding gene (INTRON, NON_SYNONYMOUS_CODING, START_LOST, STOP_GAINED, STOP_LOST, SYNONYMOUS_CODING, SYNONYMOUS_STOP, UTR_3_PRIME, UTR_5_PRIME).

Annotation of loci that were subjected to further investigation was individually updated by manual searches of the Araport 11 genome annotation (Cheng et al. 2017) accessed via the BAR Thalemine portal (Pasha et al. 2020).

To determine the direction of phenotypic parameter cHanges for individual minor allele amino acid substitutions, lists of accessions carrying specific substitutions retrieved from the Ensembl resource (Yates et al. 2022) were used to identify subsets of our accessions carrying individual sequence variants. Average values of the phenotypic parameters of question were subsequently determined for each such variant.

4.7 Geographic Mapping of Allele Distribution

Lists of accessions carrying specific alleles of selected loci for the purpose of environmentally correlated allele distribution analyses and for map generation were generated from Ensembl data as described above. Geographic coordinate-based maps were created using Python Folium (Python Visualisation 2024) or Google Maps tools.

4.8 Enrichment and Depletion Analyses of GWAS Candidate Gene Lists

To determine whether lists of candidate genes identified in the GWAS are enriched or depleted for specific gene groups, we used the Arabidopsis genome (Araport 11 version) and compared our data lists after manual annotation with the following gene lists.

As mobile element-derived genes, those with annotations containing the word ‘transposable’ were considered.

As genes with high transcript levels in trichomes, we considered the 164 genes listed as encoding 5% most abundant transcripts in the mature trichome transcriptome (Jakoby et al. 2008).

As genes with guard cell-specific transcription patterns, we considered the list of loci specifically upregulated in the guard cells compared to the mesophyll (Leonhardt et al. 2004).

As genes encoding trichome-abundant proteins, we selected genes from the recently published Arabidopsis trichome proteome study (Huebbers et al. 2022) using the following criteria. A protein had to be significantly enriched in at least two of the four reported proteomic experiments, at least in one case by a factor of two and more, while it was not depleted in any experiment. These criteria produced a list of 455 loci.

For genes encoding guard cell-abundant proteins, we considered the list from a published guard cell proteome study (Zhao et al. 2008).

For gene family enrichment, we included all gene family members as identified by keyword search of candidate gene annotations. Family member counts were based either on Literature, as in the case of F-box proteins (Kuroda et al. 2002), on TAIR (Reiser et al. 2024) Gene FamiLies annotations (as in the case of receptor-like protein kinases and FH2 proteins), or estimated by keyword searches of the Araport 11 annotations (for all DUFs).

Genes from each list were identified among the GWAS candidates and significance of any observed enrichment or depletion was evaluated using pairwise χ² test (Stagroom 2024) with Benjamini–Hochberg correction for multipLicity performed using an online calculator (Radua, Albajes-Elsagirre, and Fortea 2024).

4.9 Verification of Candidates

Loss-of-function mutants in candidate genes were grown alongside corresponding wild type plants and leaf samples were harvested, processed and imaged as described above for the main ecotype screen. To minimize effects of observer's expectation bias, a single-bLinded experimental design was employed, with one member of the team performing the imaging and others quantitatively evaluating microphotographs with coded labels without knowledge of the plant's genotype. After decoding the identity of the samples, significance of between-genotype differences in the relevant categorical parameters was estimated by the χ² test (or Fisher's exact test in cases where some categories exhibited zero counts) using online calculators (Stagroom 2024; Vasavada 2016).

4.10 Mapping of Protein–Protein Interactions Among Candidature Gene Products

To gain initial insight into the interactions among candidate genes, the full network STRING protein–protein interaction database v. 11.5 (Szklarczyk et al. 2023) was searched with non-redundant lists of genes associated with the indicated trait groups as queries, using high confidence (score = 0.7) and high stringency (FDR = 1%) settings. Statistical significance of interaction enrichment was obtained during this search. Resulting interaction networks were exported both as tables (further processed in Excel to generate cluster and node lists that were subsequently annotated as described above) and as vector images that were manually edited (to remove unlinked nodes and tidy up the layout), annotated and coloured.

Mapping of interactions among the subset of candidates associated with multiple trait groups was performed analogously except that the confidence threshold was lowered to medium (score = 0.4), that is, to the default settings of the STRING database search tool.