Systems analysis of leaf anatomy and transcriptome using unsupervised machine learning provides insight on photosynthesis development and evolution in non‐model grasses

Abstract

The C₄ photosynthesis includes intriguing leaf anatomies. The current model supports the placement of C₃‐C₄ intermediates as a middle point in the evolutionary trajectory from C₃ to C₄ photosynthesis. The known determinants involved in the differentiation of divergent photosynthetic leaves arose from the comparative analysis between both ends, C₃ and C₄ species. However, much more could be known if evolutionarily close species were analyzed together with intermediate species using advanced‐omic approaches. In the present work, by combining leaf anatomical traits and transcriptomic data with machine learning methods, we provided insights on gene regulatory networks involved in complex leaf anatomical characteristics in non‐model grasses of subtribe Otachyriinae. For that, self‐organizing maps (SOMs) were developed to group genes and phenotypic traits into clusters (neurons) according to their behavior along the leaf developmental gradient. The analysis allowed us to identify a set of genes as potential enablers of key anatomical trait differentiation related to bundle sheath (BS) cell size, vein density, and the interface between mesophyll and BS cells. At the same time, we identified genes that displaced together with the adjustment of the BS cell area suggesting a possible role in the evolution of this distinctive leaf anatomical trait.

1. INTRODUCTION

C₄ photosynthesis is a mechanism that has evolved in some plant species that results in the suppression of photorespiration. Photorespiration occurs during the carbon fixation process due to the dual activity of ribulose‐1,5‐bisphosphate carboxylase/oxygenase (RUBISCO). To increase the concentration of CO₂ around RUBISCO, most C₄ species develop two specialized leaf cell types: mesophyll cells (M) and bundle‐sheath cells (BS) (Hattersley, 1984; Sage et al., 2012). The M cells are clustered around the BS cells in a ring‐like fashion, which is a hallmark of Kranz anatomy. The BS cells house RUBISCO and have unique anatomical and genetic peculiarities like the root endodermis (Slewinski et al., 2012). In addition, leaf traits such as higher density of vascular bundles (VB), a higher fraction of the leaf occupied by BS cells, and the close contact between M and BS cells, have been identified as essential requirements for C₄ photosynthesis (Christin et al., 2013; Christin & Osborne, 2013; Edwards & Voznesenskaya, 2011; Ermakova et al., 2020; Hattersley et al., 1977; Leegood, 2002; Lundgren et al., 2019; Nelson, 2011; Sage et al., 2012; Sage et al., 2014; von Caemmerer & Furbank, 2003).

Current models indicate that the evolution of C₄ photosynthesis was a process that included the adjustment of the leaf anatomy, the relocation of photorespiratory enzymes, and the occurrence of intermediate photosynthetic subtypes (Blätke & Bräutigam, 2019; Khoshravesh et al., 2020; Mercado & Studer, 2022; Sage et al., 2014; Stata et al., 2019). Species that primarily utilize the C₃ pathway but have anatomical and physiological characteristics that reduce photorespiration are categorized as Proto‐Kranz (PK) and C₂ type I species, while others have an incipient C₄ pathway, as in the C₂ type II or the C₄‐like (Mercado & Studer, 2022; Sage et al., 2014). In terms of leaf anatomy, the PK species have traits that reduce photorespiration such as a larger area of BS with more organelles (chloroplasts, mitochondria, and larger vacuoles) arranged like that observed in C₂ species (Muhaidat et al., 2011). In C₂ species, the photorespiratory pathway is partitioned between M and BS, with the release of CO₂ occurring predominantly in BS (Rawsthorne et al., 1988). In terms of genes, overall, the known determinants involved in the differentiation of divergent photosynthetic leaves arose from the comparative analysis between both ends, model C₃ (e.g., rice and Arabidopsis) and C₄ species (e.g., maize and Setaria). For instance, transcription factors such as SCARECROW (SCR), SHORTROOT (SHR), and family members of INDETERMINANT DOMAIN (IDD), GOLDEN2 (G2), and GOLDEN2‐LIKE1 (GLK) are known to have a pivotal role in the differentiation of BS‐ and M‐cell type and vein formation (Coelho et al., 2018; Hall et al., 1998; Liu et al., 2023; Slewinski et al., 2012; Wang et al., 2017). However, much more could be known if evolutionarily close species were analyzed together with intermediate species using advanced‐omic approaches. In this sense, one of the biggest challenges of such studies is the work with non‐model species that lack high‐quality reference genomes that facilitate the mapping of omics data (reviewed in Mercado & Studer, 2022). Recently published the novo transcriptomes assemblies of Flaveria, Moricandia, Salsola, Alloteropsis and members of the grass Otachyriinae subtribe offer unique opportunities to gain additional insight into photosynthesis divergence (Dunning et al., 2019; Lauterbach et al., 2017; Mallmann et al., 2014; Prochetto et al., 2023; Schlüter & Weber, 2016).

Grass subtribe Otachyriinae is especially rich in different photosynthetic pathways including a C₄ (Anthaenantia), C₂ (Steinchisma) and PK (Rugoloa) genera and several C₃ species (Acosta et al., 2014, 2019). Recently, the first genomic analysis of species from the Otachyriinae subtribe was conducted using three perennial species of the Otachyriinae subtribe distributed in humid and tropical regions of America (C₃, Hymenachne amplexicaulis; PK, Rugoloa pilosa; and C₄, Anthaenantia lanata) (Prochetto et al., 2023). Such high‐quality leaf transcriptomes can be used as the foundation to further explore unknown and complex relationships among gene expression levels and patterns with the differentiation of phenotypic leaf traits. This implies the integration of different types of data obtained from multiple and heterogeneous sources. In this sense, self‐organizing map (SOM) has gained special attention among scientists who proved the potential of the method to solve multiple biological questions by integrating omics with phenotypic data (Nakayama et al., 2018; Kohonen et al., 2001; Kim et al., 2007; Betts et al., 2020).

A SOM is a special type of machine learning (ML) model, which has proven to be very well‐suited for the task of heterogeneous data integration and visualization through unsupervised learning (Betts et al., 2020; Milone et al., 2013; Mohnike et al., 2023; Stegmayer et al., 2012; Watanabe & Hoefgen, 2019). SOMs can represent complex high‐dimensional input patterns into a simpler low‐dimensional discrete map, with prototype vectors (neurons) that can be visualized in a two‐dimensional lattice structure and which preserve the proximity relationships of the original samples (Kohonen et al., 2001). Interestingly, SOM has several advantages over other current methods (Chai et al., 2021; DiLeo et al., 2011). Indeed, SOMs look at each data sample individually, they cluster genes with highly similar expression patterns along the features measured, and allow an individual, rather than a global, analysis of each gene/phenotype. In plant studies, SOMs have been widely used for the integration of transcriptome profiles, metabolites, the elucidation of gene‐to‐gene and metabolite‐to‐gene pathways (Allen et al., 2010; Hirai et al., 2005; Yano et al., 2006; Yokota Hirai et al., 2004), and for integration and discovery of coordinated variations in transcriptomics and metabolomics data (López et al., 2015; Milone et al., 2010; Stegmayer et al., 2009).

In this study, we carried out a system analysis of leaf transcriptome and anatomy using unsupervised machine learning to gain insight into leaf anatomy differentiation and evolution in non‐model grasses. For this, expression data from Prochetto et al. (2023) was combined with anatomical information extracted from a detailed study of the leaf gradient in four non‐model grass species from the Otachyriinae subtribe. We focused on three main components of leaf anatomy: (1) BS area, (2) patterns of venation, and (3) connection between M and BS cells. Then, anatomical data and transcriptome information were combined using unsupervised ML methods to identify a set of hidden genes in complex leaf anatomical traits of non‐model grasses.

2. MATERIALS AND METHODS

2.1. Species selection

To carry out anatomical studies, four species of the Otachyriinae subtribe were selected: H. amplexicaulis (Rudge) Nees (C₃), Rugoloa pilosa (Sw.) Zuloaga (PK), Steinchisma hians (Elliott) Nash (C₂), and A. lanata (Kunt) Benth (C₄). Species selection was based mainly on phylogenetic proximity, photosynthetic pathway, and material availability. Seeds or rhizomes of the studied species were collected in the field. The specimens were deposited in the herbaria Instituto Botánica Darwinion (SI) and Arturo Ragonese (SF). The collection vouchers are listed as follows: H. amplexicaulis (Rudge) Nees Prochetto and Reinheimer 1 (SF), Rugoloa pilosa (Sw.) Zuloaga s / n (SI), S. hians (Elliott) Nash Marino s/n (SF), and A. lanata (Kunt) Benth Acosta (SI).

2.2. Plant growth conditions and sampling

Individuals of four Otachyriinae species were grown, from seed or rhizome, in growth chambers at 27 °C under long‐day conditions (16 h of light and 8 h of darkness). For each species, the 5th young leaf (counting from the base of the plant) of 10 individuals was collected following Prochetto et al. (2023). To study the leaf development gradient, each leaf was divided into two sections, using the ligule of the 4th leaf as a marker to define the sink‐source transition zone (Li et al., 2010; Prochetto et al., 2023; Figure S1). The sections were then divided into four segments of equal length and labeled S1 to S8 from the base to the tip of the leaf. Segments 1, 3, 5, and 7 were used for the analysis. To account for the high variability between samples, replicates were paired throughout the analysis.

2.3. Preparation and analysis of samples for light microscopy

Fresh segments were placed in plugs with 5% low melting point agarose in .05 M PBS at 50 °C and left until solidification at 4 °C for at least 1 h. Cross sections of 100 μm thickness were obtained with a vibrating blade microtome (Leica VT1000 S) and mounted on slides with .05 M PBS. Sections were visualized and photographed under fluorescence microscopy using a Nikon Eclipse E200 microscope with a mercury arc lamp light source. Filter cube 96310/UV2EC (excitation 360–340 nm/emission 460–450 nm) was used to capture autofluorescence from plant cell walls. Sections were also photographed under brightfield white light microscopy (Nikon Eclipse E200). Image processing and anatomical measurements were performed using FIJI v2.9.0 (Schindelin et al., 2012). Phenotypic traits were measured along the entire leaf cross section. Definitions of phenotypic traits and units are shown in Table 1. The determination of cell interface parameters (Sv* and Sb*) is shown in detail in Figure S2.

TABLE 1.

Definition of quantitative phenotypic traits.

Phenotypic trait	Definition	Units
VB distance	Distance between central VB and its closest VB	[μm]
VB density	Number of VB per cross section area	[mm2]−1
1° VB density	Number of 1° VB per cross section area	[mm2]−1
2° VB density	Number of 2° VB per cross section area	[mm2]−1
1° VB/Total VB	Proportion of 1°VB	‐
VB diameter	Mean VB diameter	[μm]
VB area %	Proportion of cross section area occupied by VB	%
IBS cell size	Mean cell area in the cross section	[μm2]
IBS area %	Proportion of cross section area occupied by IBS	%
OBS cell size	Mean cell area in the cross section	[μm2]
OBS area %	Proportion of cross section area occupied by OBS	%
M cell size	Mean cell area in the cross section	[μm2]
Sb*	Cell interface between BS and M per cross section area	[mm]−1
Sv*	Cell interface between VB and BS per cross section area	[mm]−1
Leaf size	Total cross section area	[μm2]

Note: A detailed description of Sb* and Sv* traits is presented in Figure S2.

Abbreviations: M, mesophyll; IBS, inner bundle sheath cells; OBS, outer bundle sheath cells; VB, vascular bundle.

2.4. Statistical analysis

To determine the significance of the differences in the phenotypic characters, non‐parametric methods were used contemplating paired samples (for intra‐species comparisons) and unpaired (for comparisons between species) with Graphpad Prism (v6.0.1 for Windows, www.graphpad.com). For the paired samples, Friedman test and Dunn's multiple comparisons test were used. For unpaired samples, Kruskal‐Wallis test and Dunn's multiple comparisons test were used. p Values lower than .05 were considered significant.

Principal component analysis (PCA) was conducted in R, using stats package (v3.6.2, R Core Team, 2016). Biplot graphs were generated using factoextra (v1.07, Kassambara & Mundt, 2022).

2.5. Gene expression data

Transcriptomic data including gene annotation and transcripts abundance was retrieved from a previous study (Prochetto et al., 2023). Gene Expression Matrices were built using the abundance_estimates_to_matrix.pl script from Trinity package v.2.8.5 (Haas et al., 2013), to generate a TPM expression matrix. Sequencing depth normalization was performed using the trimmed mean of M‐values (TMM) method from the EdgerR package v. 3.38.1 (Robinson et al., 2009). Genes between species were analyzed using orthologs inferred from Orthofinder (Emms & Kelly, 2019). In most cases, orthogroups were made of single copy orthologs. In the case of multicopy orthogroups, the sum of the expression values of the transcripts was used. The total number of genes used in this study was 9,739.

2.6. Global SOM

2.6.1. Size optimization

In this work, we have developed two SOMs for different purposes. The first global SOM model was developed for grouping features (that is, genes together with phenotypic traits) into neurons according to their expression along leaf development and species with different photosynthetic subtypes. Before building the global SOM, the optimum grid size was studied in the following way. A median neuron size of 32 to 37 features per neuron was set (Table S1) to make the analysis of the results of the individual neurons affordable. The total number of features (genes and phenotypic traits) measured for all three species was 13,953. By using different cutoff expression values as a threshold for the genes, several subsets were obtained (ranging from 13,953 to 4206 according to the median number of features per neuron desired) and used to build SOMs of different sizes (from 400 to 121 neurons in total). For each SOM, its corresponding quantization error, a parameter that assesses the accuracy of the SOM for representing the data, was calculated with the somQuality function from the aweSOM package (v1.3, Boelaert et al., 2022). The minimum quantization error was achieved for the SOM with 289 neurons (17x17 map) with a median number of features per neuron of 32 (Table S1). The selected SOM includes 9739 genes, 15 phenotypic traits and three species‐specific features (Dataset S1). The scaled expression values (log transformed, mean centered and variance‐scaled) of a total of 9757 features for all the three species under analysis were used. SOM training was made with aweSOM package (v 1.3., Boelaert et al., 2022), using a hexagonal grid of 289 neurons (17x17), rlen = 1000, alpha = c (.1, .001) and Euclidean distance function.

2.6.2. SOM validation

To evaluate the robustness of neurons enriched for specific phenotypic traits, we performed a perturbation analysis in the following way. First, we obtained the distance from the selected phenotypic traits to the corresponding neuron centroid. Second, we systematically removed those specific genes (that were clustered with selected phenotypic traits) and re‐calculated the distance from the phenotypic traits of interest to the new corresponding neuron centroid. After this perturbation, the two possible results are the following: (i) the distance of each phenotypic trait of interest to the new neuron centroid does not change, indicating no true relationship among the genes removed and the phenotypic traits; (ii) the distance of each phenotypic trait of interest to the new neuron centroid is increased (worsened), indicating a true and cohesive relationship between the genes removed and the phenotypic traits patterns (Table S2).

2.7. Species‐specific SOM

The second model was a species‐specific SOM built in order to focus on expression and phenotypic patterns of variation instead of their absolute expression magnitude; and to identify features that vary between species. Expression values were mean centered and variance‐scaled separately for each species. Then, C₃ data (only) was used to train a 3 × 2 hexagonal SOM. This small size for the map was chosen for limited redundancy in co‐expression patterns across clusters as in (Betts et al., 2020), that is, in order to have a clear and easy way to analyze patterns of variation along leaf development in each neuron, as suggested in (Nakayama et al., 2018). Later, data from PK and C₄ species was mapped to the species‐specific SOM trained with C₃ data. To visualize (as a directed network) the assignment of features from different species to separate neurons, igraph package v1.3.5 (Csárdi & Nepusz, 2006) was used. Clustered and displaced feature sets among clusters were subjected to gene ontology analysis. To validate the displacements, we conducted a SOM perturbation analysis, similar to the one performed for the global SOM.

2.8. Gene Ontology enrichment analysis

The Gene Ontology (GO) enrichment analysis of each cluster was performed using the R package TopGO v2.42 (Alexa & Rahnenfuhrer, 2020). GO terms for each transcript were obtained from Prochetto et al. (2023) annotation matrices. The runTest() function was used to test for enrichment using Fisher's exact test. p Values were adjusted using the “elim” algorithm. A term was significant if its adjusted p value < .01. To characterize neurons, GO BP terms were extracted and summarized with REVIGO tool (http://revigo.irb.hr, Supek et al., 2011).

3. RESULTS

3.1. Leaf anatomy characterization

To supplement available anatomical information on Otachyriinae species (Khoshravesh et al., 2016; Lundgren et al., 2014), we analyzed the anatomy and development of S. hians (C₂) leaf and compared them with additional anatomical observations in H. amplexicaulis (C₃), R. pilosa (PK) and A. lanata (C₄) (Figure 1; Figure S3; Prochetto et al., 2023). Overall (Figure 1; Figure S3), in the leaf cross sections of S. hians, H. amplexicaulis, and R. pilosa, we observed that VB are arranged linearly; each of the VB is surrounded by two layers of cells: the smallest forms the inner bundle sheath (IBS) and the largest forms the outer bundle sheath (OBS). In these cases, four or more M cells are placed in between VB. In contrast, the C₄ leaf cross sections of A. lanata showed a three‐dimensional venation system as previously described for some thick C₄ leaves (Lundgren et al., 2014; Ocampo et al., 2013); each of the VB is surrounded by a single layer of bundle sheath cells (IBS). In this case, two to three M cells are placed in between VB.

FIGURE 1.

Developmental gradient in the 5th leaf of four Otachyriinae subtribe species. (A) S1 (leaf base) and S7 (leaf tip) cross section cuts from C₄ A. lanata , C₂ S. hians , PK R. pilosa, and C₃ H. amplexicaulis . (B) Magnification showing 1° vascular bundles from S7. Abbreviations: 1°VB, primary vascular bundle; 2°VB, secondary vascular bundle; IBS, inner bundle sheath cell; M, mesophyll cell; OBS, outer bundle sheath cell.

In order to gain a comprehensive understanding of the developmental differences among species, we selected 15 leaf anatomical traits that were measured along the four selected segments to expand the leaf developmental gradient (Table 1; Figure S4).

3.1.1. Comparison of traits associated with VB system, BS fraction, and M/BS contact surface between C₄ and non‐C₄ species

Leaf traits such as higher density of vascular bundles (VB), a higher fraction of the leaf occupied by BS cells, and the close contact between M and BS cells, have been considered essential requirements for C₄ photosynthesis. Here, we found that C₄ species exhibited the shortest distance between VB, resulting in a higher VB density. Overall, an increasing VB density was observed from the C₃ species to PK species, and then to C₂ and C₄ (Figure 2A). The proportion of primary VB over the total of VB was significantly smaller in C₄ (Figure 2A). Leaf fraction occupied by IBS was significantly higher in C₄ (Figure 2B). Similarly, the leaf fraction occupied by OBS was significantly higher in C₂ and lower in PK while this cell type is absent in the C₄ species (Figure 2B).

FIGURE 2.

Phenotypic traits comparison between species in S5 (A, B, C) and S1 (D, E, F). (A, D) Phenotypic traits associated with vascular bundles. (B, E) Phenotypic traits associated with bundle sheath cells. (C, F) Parameters associated with interfaces between cell types. Asterisks point out statistical differences between groups, lines otherwise (Kruskal‐Wallis test and Dunn's multiple comparisons test, p value < .05). Abbreviations: Al, Anthaenantia lanata ; Ha, Hymenachne amplexicaulis ; IBS, inner bundle sheath cells; M, mesophyll cells; OBS, outer bundle sheath cells; Rp, Rugoloa pilosa; Sh, Steinchisma hians ; VB, vascular bundle.

The contact surface parameter S_b, defined by Pengelly et al. (2010), is calculated on leaf cross section and by measuring the perimeter of the BS within the space between two VB and dividing it by the distance between VB. However, this parameter does not consider the three‐dimensional organization of the VB in A. lanata. So, to include the tridimensional organization of VB in Pengelly's equation, we here proposed a variation of this parameter (S_b*). The S_b* parameter is calculated as the perimeter of all the BS cells in contact with the mesophyll divided by the cross‐sectional area of the leaf. The results showed that C₂ and C₄ S_b* are significantly larger than the C₃ species (Figure 2C). We also estimated the contact surface between BS and VB (S_v*), observing that C₂ and C₄ had significantly larger values than C₃ species (Figure 2C).

To assess when these anatomical differences occur along the leaf development, we quantified the same traits in immature leaf segments (S1) (Figure 2D–F). Overall, these measurements indicate that differences between C₄ and non‐C₄ species, although established early in development, tend to increase significantly along the leaf gradient.

3.1.2. Discovering leaf developmental diversity through PCA analysis

To integrate the findings, we conducted a PCA of the samples using 15 phenotypic traits (recovered from S1 to S7) as variables (Figure 3). The PCA distinguished between C₄ and the rest of the species and showed a partial differentiation between PK and C₂ species (Figure 3A). Principal component 1 explains 48.8% of the variation and principal component 2 explains 28.8% of the variation. Phenotypic traits, illustrated as vectors in the biplot, revealed that M cell area, IBS area, and to a lesser extent, VB area %, were influential in distinguishing C₄ species from the rest. On the other hand, traits related to VB density and contact areas (S_b*; S_v*) are positively correlated with the shift from C₃ to C₂ (Figure 3A). To confirm these findings, we calculated trait contributions to PC1 and PC2 (Figure 3B). VB‐associated traits like Sv*, VB density, and 2 VB density were prominent contributors to PC1, explaining 55.3% of the variation. In contrast, PC2 was dominated by BS‐related traits, including M cell size, IBS cell size, and OBS Area %, collectively contributing 50.6% of the variation. This underscored the distinctive contributions of VB and BS traits to the observed developmental differences.

FIGURE 3.

Summary of leaf development traits. (A) Principal component analysis of the samples using phenotypic traits (PCA biplot). (B) Contribution of variables to PC1 and PC2. Phenotypic traits associated with vascular bundle and bundle sheath cells are colored in pink and salmon, respectively. (C) Development of key phenotypic traits along the leaf gradient. Mean values for C₃ (full light green circles), PK (empty green circles), C₂ (full dark green squares), and C₄(empty light blue squares). Abbreviations: 1°VB, primary vascular bundle; 2°VB secondary vascular bundle; IBS inner bundle sheath cell; M mesophyll cell; OBS outer bundle sheath cell; PC, principal component; Sb*, cell interface between bundle sheath cell and mesophyll cell per cross section area; Sv*, cell interface between vascular bundle and bundle sheath cell per cross section area.

It is well documented that an increase in the BS size and their arrangement around VB are among the most relevant characteristics of the Kranz anatomy (Esau, 1953). In line with this, we found that the IBS area % was up to 4 times higher in the C₄ species compared to non‐C₄ species (Figure 2B). The increase in BS area % in the C₄ leaf may be attributed to both larger IBS cell sizes surrounding primary VB and additional IBS cells around several secondary VB (Figure 2B; Figure S4). In addition, in the C₄, IBS cell size is significantly higher from S1 and keeps increasing as the leaf matures, while in non‐C₄ species the trait appears to be established early in the leaf gradient (Figure 3C, Figure S4).

The distance between VB tends to increase throughout the leaf development in C₃, PK, and C₂ species while decreasing in C₄ species (Figure S4). This trend leads to a decrease in VB density as leaves mature in PK and C₂ but an increase in VB density in C₄. The pattern in C₃ species is less clear, with VB distance increasing in S3 and then remaining relatively unchanged (Figure S4). We observed that significant changes in VB distance and density occur primarily between leaf segments S1 and S3 in all studied species (Figure S4). Further analysis reveals that the formation of secondary VB is the key factor driving the differences in VB density between species, particularly evident in the substantial increase in secondary VB numbers between segments S1 and S3 (Figure 3C; Figure S3).

Given the relevance of the VB‐ and BS‐related traits in differentiating leaf anatomies in Otachyriinae, we decided to explore the genetic basis underlying the observed variations with a focus on secondary VB formation, IBS cell size and the interface between BS and M cells. For that, we propose here a novel machine learning methodology for discovering hidden genes in complex leaf anatomical traits in non‐model Otachyriinae grass species.

3.2. Combining phenotypic traits with gene expression data

To identify subsets of genes with similar expression profiles and link them with the behavior of phenotypic traits along the leaf gradient we built a global SOM. For this, the Orthogroups (OG) expression from three species (A. lanata, R. pilosa, and H. amplexicaulis) and four segments of the leaf (S1, S3, S5, and S7) (Prochetto et al., 2023) were merged with phenotypic trait data. This SOM includes 9739 genes, 15 phenotypic traits, and three species‐specific features (Figure 4, Dataset S1). To facilitate the analysis of grouped features at the neuron level, an interactive version of the SOM is provided in Figure S5. For exploratory and visualization purposes, phenotypic traits are highlighted alongside genes associated with leaf development processes or relevant cell functions like cell walls, plastids, photosynthesis, plasmodesmata, and transcription factors. Other features were categorized as “other features.” Overall, genes involved in the development of cell walls, plasmodesmata and plastids, as well as transcription factors are widely distributed along the map. Meanwhile, some classes like photosynthesis and phenotypic traits are concentrated in more specific areas (left and central bottom area for photosynthesis, central and right bottom area for phenotypic traits). When the average values of the grouped features were analyzed, we observed, as expected, that close neurons present similar patterns (average expression value of all the features grouped in the neuron) along the samples (Figure S6). We also looked at transcription factors that were previously described as important for M and BS cell differentiation in rice and maize: SCARECROW (SCR), SHORTROOT (SHR), and GOLDEN‐LIKE (GLK) (Hall et al., 1998; Slewinski et al., 2012; Wang et al., 2017). Of them, only SCR was included in the SOM. We found SCARECROW (OG0001885) in neuron 117 (Figure S5). Neuron 117 is composed of genes mostly related to protein binding and transport. The SHR and IDD orthogroups were not included in the analysis because they were filtered by low expression level. In addition, the GLK orthogroup was not involved in the SOM given the absence of the C₄ GLK counterpart.

FIGURE 4.

Self‐organizing map from Dataset S1 (with a total of 9757 features coming from leaf phenotypic information (18 features) and gene expression data (9739 genes) for C₄ A. lanata , PK R. pilosa, and C₃ H. amplexicaulis . Each cell (hexagon) represents a neuron and each gray point a feature. Neurons containing phenotypic traits are highlighted in light green. To visually explore feature allocation patterns, we highlighted specific features in colors based on their functional roles: cell wall (red), plastids (orange), species (yellow), phenotypic trait (green), photosynthesis (blue), plasmodesmata (cyan), transcription factor (purple) and other (gray). The neurons that are discussed in detail in the manuscript (N14, N65, N161, N228) are indicated with its corresponding number and have a thicker black contour.

3.2.1. GO terms in selected neuron pairs

Given that SOM only considers positive relations among features, we looked for pairs of neurons (clusters) with opposite patterns in order to find neurons that better represent both types (positive and negative) of potential relationships between genes and phenotypic traits. Pairs of neurons having phenotypic traits of interest and opposite patterns associated with GO terms are shown in Figure 5. A detailed membership list of each neuron pair is shown in Dataset S2.

FIGURE 5.

Selected pairs of opposing neurons containing phenotypic features: N14 and N161 (A, B), and N65 and N228 (C, D). Neuron median expression values along samples (A, C, E). Semantic similarity clustering of biological processes gene ontology terms (GO) present in neuron pairs (B, D). Circle colors indicate if a term is present in one neuron (red and blue) or in both neurons in the pair (gray). Circle size indicates the times in logarithmic scale that a term is present in the reference database. Abbreviations: 1°VB, primary vascular bundle; 2°VB, secondary vascular bundle; IBS, inner bundle sheath cell; M, mesophyll cell; PhT, phenotypic trait; Sb*, cell interface between bundle sheath cell and mesophyll cell per cross section area.

The IBS cell area and M cell area were placed in neuron N14 from the N14‐N161 pair (Figure 5A). N14 contains features with high expression in C₄ and low expression in PK, with C₃ in between. A wide range of processes represented as GO terms clusters were present in the pair (Figure 5B), including defense response, protein ubiquitination, plant ovule development and regulation of secondary growth as the most numerous clusters. Although most of the terms were present only in one of the neuron pairs, most of the clusters presented terms from both neurons (Dataset S2). This pair also contains one transcription factor (OG0011232, inside N161) belonging to the NF‐YA family.

To validate the findings and as a quality metric of the clusters, we extracted distances from features to centroids from the global SOM object, where shorter distances indicate a more robust correlation between feature and neuron expression patterns. To facilitate cross‐neuron comparisons, we globally scaled distances using Z‐scores. Negative values in scaled distances indicate features very close to the centroid or the “core” of the neuron, while high and positive values suggest features located in the “periphery” of the neuron. Specifically, scaled distances for IBS cell area and M cell area are −1.30737 and −.4759, respectively. These values indicate that both features reside in the core of N14, showing that this neuron accurately represents the expression pattern of these features of interest. On the other hand, the scaled distance for TF OG0011232 to N161 is negative (−.02977371), affirming that this transcription factor is a representative feature of the neuron.

VB density, 2° VB density and S_b* phenotypic traits were assigned to N65 in the global SOM, belonging to the N65‐N228 pair (Figure 5C). N65 contains features with higher expression in C₄ than PK and C₃ species, while neuron 228 shows the inverse pattern. GO terms present in the pair were clustered in 13 clusters (Figure 5D; Dataset S2), with seven of them being single term clusters. The remaining six clusters were involved in response to salt stress, transcription and regulation of transcription, cell differentiation, proton transmembrane transport and abscisic acid catabolic process. Scaled distances for VB density and 2° VB density indicate that both phenotypic traits reside in the core of N65 (−.45657 and −1.01809 respectively). In contrast, S_b* scaled distance (1.556816) suggests that the neuron does not faithfully represent the expression pattern of this phenotypic trait, limiting the connection between S_b* and the other features allocated in the neuron. This pair also included five transcription factors, but only three of them had negative scaled distances: one inside N65 (OG0005274, −.23199254) and two inside N228 (OG0001295, −.89068752; and OG0004381, −.91866897).

3.2.2. Global SOM validation

To validate the relationships found in the global SOM, we performed a mathematical perturbation analysis by systematically removing the genes clustered with selected phenotypic traits, as explained in Methods. We found that the scaled distances to the selected phenotypic traits centroid actually increased after the removal of genes from the dataset. This observation suggests that the removed features have the strongest relationships with the selected phenotypic trait among all genes in the dataset (Table S2). Additionally, we employed bibliographic data for further confirmation of the groups found. We looked for transcription factors with well‐established roles in Zea mays and Oryza sativa leaves (Cai et al., 2017; Dou et al., 2021; Wu et al., 2015; Xiang et al., 2021). After confirming the presence of orthologs in our dataset, we identified the neuron pairs containing them and the genes clustered there. For those genes, we performed a GO analysis with REVIGO. We found that for the Z. mays and O. sativa transcription factors with known involvement in drought stress, defense response, and hormone signaling, the corresponding orthologs in our dataset were clustered in neurons having genes associated with similar GO terms (Dataset S3).

3.2.3. Species‐specific SOM for studying conservation of expression and phenotypic patterns of variation across evolution

To study the changes in features (genes and phenotypic traits) expression patterns associated with photosynthetic evolution, a species‐specific SOM was built with the C₃ data alone. The grid size was chosen to be small (3 × 2) so that neurons represent unique and non‐redundant expression patterns which can capture the behavior of features along the leaf gradient of our dataset (Figure S7). Then we used the C₃ species‐specific trained SOM to map PK and C₄ data to identify two different scenarios: (a) features that are preserved in the same neurons among species and (b) features that are allocated in different neurons among species. Indeed, a feature with a very similar expression pattern across different species will be always mapped to the same neuron in the trained map, meaning that the feature is preserved across all species in this study. In contrast, a feature mapped to a different neuron means that the feature expression pattern is different between species, and therefore, the feature will be displaced (from a particular neuron in species A to another neuron in species B). The analysis performed here assumes that feature preservation status gives information on the preservation of gene regulatory networks between species. To summarize the results, we have built a network graph to show feature allocation to neurons and displacements from C₃ to PK (Figure 6A) and from PK to C₄ (Figure 6B). Furthermore, the quality of the displacement was assessed by inferring the relationships between the neuron expression patterns. Based on the results we identified four groups of expression pattern displacements: early, when it looks shifted to more immature segments in the leaf gradient; delay, when the displacement looks shifted to more mature segments in the leaf gradient; flip, when it presents the opposite pattern; and others, that groups the rest of expression pattern displacements with no specific meaning.

FIGURE 6.

Displacement of features in different clusters in the SOM clustering scheme. Network representations of feature assignment into different SOM clusters using the displaced features. Arrows represent displacement from C₃ to PK (A) and from PK to C₄ (B). Arrows and circle sizes are proportional to the number of displaced features and neuron population, respectively. Line plots indicate representative expression patterns (neuron median expression) in each cluster throughout the leaf gradient. Arrow colors show distinct displacement types (delay, flip, early, other). Abbreviations: 1°VB, primary vascular bundle; 2°VB, secondary vascular bundle; IBS, inner bundle sheath cell; M, mesophyll cell; OBS, outer bundle sheath cell; Sb*, cell interface between bundle sheath cell and mesophyll cell per cross section area; Sv*, cell interface between vascular bundle and bundle sheath cell per cross section area; VB, vascular bundle.

A total of 5469 features were displaced between neurons from C₃ to PK (Figure 6A, Figure S8), and the remaining 4280 features were allocated to the same neuron in both species (non‐displaced features). Almost 70% of displaced features were displaced to a close neuron (3786 features), either to an earlier (1145 features, 20.9%) or delayed (2641, 47.7%) expression pattern. Only 438 features (8.0%) presented a flip displacement, while other kinds of displacements altogether accounted for 1245 features (22.7%). Just three of the phenotypic traits were preserved between C₃ and PK (IBS cell area, VB area %, and 1°VB/Total VB). The early group of displaced features contains one phenotypic trait (VB diameter) and is enriched in carboxylic acid catabolic process, response to stress, and leaf development, among others (Dataset S4). Two phenotypic traits belong to the delay group of displaced features (leaf area and OBS cell area). A GO enrichment analysis showed that this group is significantly enriched in RNA metabolic processes (Dataset S4).

A total of 5474 features were displaced from PK to C₄, while 4271 were allocated to the same neuron in both species (non‐displaced features). Most of the features were displaced to a close neuron (67.7%) as in the C₃ to PK transition (Figure 6B, Figure S8). However, these account for a higher number of early displacements (1687, 30.4%) and a lower number of delay displacements (2068, 37.3%) than in the C₃ to PK transition. Also, while the delay displacements show a homogenous distribution across neurons between C₃ and PK, in the PK to C₄ transition, some displacements, such as from neuron 1 to 3 or neuron 2 to 5 are rare. A higher number of phenotypic traits (leaf area, VB area %, 1°VB/Total VB, VB diameter, 1° VB density, and S_v*) were preserved in this transition than in the C₃‐PK transition. A GO enrichment analysis showed that the early group of displaced features is enriched in mRNA processing and seed germination, among others (Dataset S4). The delay features from PK to C₄ are significantly enriched in developmental processes involved in reproduction, proteolysis, cell morphogenesis, and developmental cell growth among others (Dataset S4).

3.2.4. Genes displaced together with specific phenotypic traits

Addressing the displacement of phenotypic traits between species is another way of identifying potential regulators involved in photosynthetic pathway evolution. Overall, there is not a clear trend in the type of phenotypic trait displacements (Figure S9). The area of the IBS is an interesting trait postulated as one of the most important in Kranz anatomy because it is where CO₂ carboxylation takes place in C₄ species (Christin et al., 2013; Lauterbach et al., 2019). Our results show that the IBS cell area is preserved between C₃ and PK species (Figure 6A, both belonging to neuron 1) but it is displaced to neuron 6 in C₄ species (Figure 6B). The IBS expression changes from one descending pattern from base to tip, to an ascendant pattern in the C₄. We found that together with this phenotypic trait there are also 41 genes displaced. A GO enrichment analysis showed that this group of genes is significantly enriched in suberin biosynthetic process, pyruvate transport and cell wall modification, among others (Dataset S5). Indeed, two genes in this group are known to be part of suberin biosynthesis; OG0000296, an ortholog to AtFAR1/4 (fatty acid reductase) and OG0005527, an ortholog to AtABCG16, a class of ABCG half‐transporters. In addition, there are two transcription factors among the features displaced. One is a member of the class 1 TCP transcription factor family (OG0009002, AtTCP7). The second one (OG0012289, AtBFP4) is a member of the GeBP family. To validate these findings, we conducted a SOM perturbation analysis, similar to the one performed for the global SOM, removing the 41 genes displaced with IBS cell size. Results did not reveal any new genes sharing the same displacement pattern.

4. DISCUSSION

The origin of C₄ photosynthesis from a C₃ ancestor, requiring changes in leaf anatomy, cell structure, biochemistry, and physiology, appears to involve dozens of genes. Complicating this scenario, C₄ photosynthesis is one of the most notable examples of evolutionary convergence, since at least 70 independent origins of it are known, grouped into three large phylogenetic groups: Poaceae (grasses), Cyperaceae, and Caryophyllaceae (Sage, 2004). In particular, at least 22 independent origins were suggested for the grasses (Grass Phylogeny Working Group II, 2012; Huang et al., 2022). Among grasses, the diverse photosynthetic groups represented by the genus Neurachne, individuals of Allopeteropsis semialata and species of the subtribe Otachyriinae have been pointed out by scientists as interesting lineages to answer multiple questions on the evolution of photosynthesis. Of these, the studies on the genus Neurachne and the special case of photosynthesis diversity analyzed from different A. semialata populations arose exceptional knowledge on the evolution of photosynthesis in grasses (Christin et al., 2013; Dunning et al., 2017, 2019; Lundgren et al., 2019; Khoshravesh et al., 2020; Pereira et al., 2023). Overall, in both grass lineages, the proliferation of minor veins and the increase in vein density were demonstrated to be a relevant trait for C₄ evolvability; however, evidence suggested that this may not be true for eudicots (Freitag & Kadereit, 2014; Khoshravesh et al., 2020; Lundgren et al., 2019; Voznesenskaya et al., 2013). In contrast, traits such as BS enlargement, which was originally thought of as a key determinant in the evolution of C₄ leaves, may represent secondary adaptation in grasses. Recently, comparisons of C₃ and C₄ genomes of A. semialata showed that they are highly syntenic and underwent few gene duplication and translocation events among the different photosynthetic groups (Pereira et al., 2023). This implies that most of the changes involved in evolution arose from the change in expression patterns and networks of the existing genetic background.

To date, the knowledge we have about the Otachyriinae subtribe is fragmented and remains behind other grass groups (Acosta et al., 2014, 2019; Lundgren et al., 2014; Khoshravesh et al., 2016; Pereira et al., 2023). In the present work, by combining anatomical traits and transcriptomic data with machine learning methods, we provided insights on gene regulatory networks involved in complex leaf development traits and photosynthesis evolution of non‐model grasses of subtribe Otachyriinae.

In the first part of the study, we measured and compared anatomical traits along the leaf of four members of the Otachyriinae subtribe that exhibit distinct photosynthesis physiology. The analysis included traits that have been demonstrated to be the essential requirement for C₄ physiology as the presence of M and BS cells to house different enzyme reactions, close contact of both cell types to promote efficient exchange of metabolites and the fractions of the leaf occupied by BS cells required to accommodate more organelles (Edwards & Voznesenskaya, 2011; Hattersley et al., 1977; Leegood, 2002; Nelson, 2011; von Caemmerer & Furbank, 2003). In addition, based on Neurachne and A. semialata works, we included several parameters to describe the role of the vein system on the different leaf types (Christin et al., 2013; Dunning et al., 2017; Khoshravesh et al., 2020; Lundgren et al., 2019).

The study of the leaf anatomy presented in this work demonstrated the existence of a leaf development gradient as was widely documented for grasses (Li et al., 2010; Majeran et al., 2010; Pick et al., 2011; Prochetto et al., 2023; Studer et al., 2016; Wang et al., 2014). In general, the most significant phenotypic changes occurred early in development (between S1 and S3) (Figure S4). Overall, VB‐ and BS‐related traits were prominent contributors that explain the variability of the sampling. The PK and C₂ leaves were most similar in terms of leaf anatomy; however, as expected, the C₂ leaves presented a higher degree of Kranz characteristics such as density of the VB, fraction occupied by the OBS and the contact surfaces between the different cell types (Figures 2 and 3). These results agree with previous anatomical observations on S. hians as well as other C₂ grass species (Khoshravesh et al., 2016).

The importance of the size of the BS as a Kranz trait is given by the assumption that the BS must be large enough to house many chloroplasts that perform the CBB cycle and thus maintain a good photosynthetic efficiency (Sage, 2004). Indeed, it has been suggested that a higher fraction occupied by BS in C₄ species compared to C₃ species is one of the distinguishing features of C₄ anatomy (Christin et al., 2013; Khoshravesh et al., 2020; Lauterbach et al., 2019; Sage, 2004). In this work, we found that in the non‐C₄ species, the BS that houses the CBB cycle (OBS) is 2 to 8 times larger and occupies more leaf fraction than the homologs (IBS) of the C₄ (Figure 2). Similar observations were reported previously for other grass species (Christin et al., 2013; Khoshravesh et al., 2016; Lauterbach et al., 2019). These results suggest that the leaf fraction occupied by the BS may not be a constraint for the photosynthesis performance of some of the plant groups such as Otachyriinae. In line with this, other traits such as the contact between the M and BS may be more relevant in this context as explained below. Interestingly, some of the C₄ species have two Kranz compartments, OBS and IBS; however, most of the C₄ species have only one. In this work, anatomical comparisons were made using a C₄ species of the NADP‐ME sub‐type that has IBS. To date, there is not enough evidence to identify the moment in which natural selection opted for IBS or OBS as the Kranz compartment. Concerning this, it is thought that C₄ species that use IBS as a Kranz compartment come from a C₄ ancestor with OBS (Christin et al., 2013).

Recently, the importance of the transport of C₄ acids through plasmodesmata between M and BS has been highlighted as a constraint for this process (Danila et al., 2016; Danila et al., 2018; Khoshravesh et al., 2020). Indeed, it has been shown that the C₄ species have a higher density of plasmodesmata at the interfaces between BS and M, as well as a greater contact surface between both cell types (Danila et al., 2016). In line with this, in Otachyriinae, we observed an ascending pattern in the extent of the interface between BS and M cells from C₃ to C₄ species (Figure 2; Figure S4). Given the lack of information on plasmodesmata density in intermediate species here presented, the place of this event in the evolution of C₄ photosynthesis in Otachyriinae is unknown. Remarkably, S. hians and A. lanata manage to increase the contact surface through different strategies (Figure 2; Figure S4). Although S. hians presents OBS with an area like those of R. pilosa and H. amplexicaulis, its high contact surface is explained by an increase in the density of VB and a decrease in the number of M between the VB. Comparing S. hians with A. lanata we see that the latter has IBS that are less than half the area of the OBS of the former. However, A. lanata has a higher density of VB, in particular secondary VB, achieving a similar contact surface with lots of minor VB. A similar trend was observed in Flaveria and grass species that use IBS as a Kranz compartment (Christin et al., 2013; Khoshravesh et al., 2020; Lundgren et al., 2019; McKown & Dengler, 2009; McKown & Dengler, 2010). The determination of the densities and areas of the plasmodesmata in these species would help to corroborate the dynamic of the interface between M and BS cells in Otachyriinae.

In the second part of the study, we proposed to crossover the leaf anatomy phenotypes described in this work with transcriptomic data previously reported (Prochetto et al., 2023) to gain insight into changes in genetic networks related to complex anatomical traits presented. For that, we built different SOMs. The goal of SOM (Kohonen, 1982; Wehrens & Buydens, 2007) is to represent complex high‐dimensional inputs (features with many characteristics or conditions) into a simpler 2D neuron structure while preserving the proximity relationships of the original data in the 2D map (Milone et al., 2010; Stegmayer et al., 2009). Essentially, SOM is a pattern‐based clustering method for grouping features into clusters represented by neurons (Mohnike et al., 2023).

Taking advantage of this, we propose to use SOM in a novel way for integrating plant transcriptome features with leaf anatomy features. Its advantages over other clustering methods lie in its versatility to handle diverse data types (e.g., transcriptomic and phenotype) and its capacity to individually analyze each data sample condition by condition. This individual analysis involves identifying patterns of individual behavior that are similar among features, forming clusters. SOM has the capability of grouping features with highly similar expression along all measured conditions. Moreover, the robustness of their clusters can be measured and quantified with the neuron's quantization error. In contrast to methods like WGCNA (weighted correlation network analysis), where samples are globally compared (not condition by condition) with a single and global correlation coefficient, SOM acts as a detailed “zoom‐in” on each feature. This approach enables the more accurate grouping of samples with other genes/phenotypic traits exhibiting similar individual behavior. For instance, in WGCNA, once a single global correlation value is obtained for each pair of features, a clustering algorithm (e.g., hierarchical clustering or k‐means) is applied to generate modules of genes that correlate globally. WGCNA obtains large modules of highly correlated genes related to phenotypes, but the analysis is conducted globally, with each phenotype examined only once concerning a module of several genes compacted together and represented by a single correlation coefficient. In contrast, SOM performs a pairwise analysis condition by condition for each feature. This deep, detailed, and individualized analysis afforded by SOM allows for finding more robust and accurate associations between individual gene behavior and phenotype behavior across conditions (Watanabe & Hoefgen, 2019).

Interestingly, the application of the global SOM allowed us to pinpoint genes and transcription factors that exhibited a direct relation with a phenotypic trait throughout leaf development, potentially playing a pivotal role in the evolution of leaf anatomy in Otachyriinae (Figure 4; Figure S4). A comprehensive list of potential gene drivers extracted from Figures 5 and 6 is provided in Table 2. It is intriguing to highlight that many of the identified genes are mentioned for the first time in the context of photosynthesis development and evolution. These results may be attributed to our close study of related grass species, contrasting with the literature that often identifies candidate genes by comparing distant grasses like rice versus maize. Furthermore, unique and intrinsic patterns of Otachyriinae leaf development, such as the three‐dimensional disposition of VB observed exclusively in A. lanata, may contribute to these novel findings. In addition, identifying gene networks in non‐model grasses, usually with different ploidy levels, represents a big challenge nowadays due to the absence of a reference genome for transcriptome assembly and gene annotation. This contrasts with the well‐established pipeline available for model species like rice and maize. Also, whole‐leaf transcriptome analyses represent the combined expression from M and BS cells and may not capture subtle differences in expression patterns of important transcription factors such as SHR, IDD and GLK that preferentially express in one type of cell. Undoubtedly, cell type–specific transcriptomics in Otachyriinae would add resolution to the analysis presented in this work.

TABLE 2.

Transcription factors (TF) potentially involved in leaf phenotypic traits differentiation.

Gene	TF family	Found in	Potential regulation	Orthologs known functions	References
OG0011232	NF‐YA	SOM 17x17	IBS cell area M cell area	Several roles in development, abiotic response	Edwards et al., 1998; Siefers et al., 2009; Yang et al., 2017; Mu et al., 2013; Yu et al., 2021; Hwang et al., 2019; Ballif et al., 2011; Warpeha et al., 2007; Li et al., 2021; Tokutsu et al., 2019; Alam et al., 2015; Lee et al., 2015; Li et al., 2017; Wang et al., 2018; Lv et al., 2022; Tan et al., 2022; Su et al., 2018.
OG0005274	GRAS	SOM 17x17	VB density 2° VB density Sb*	Shoot and root indeterminacy	Schulze et al., 2010
OG0001295	CAMTA			Balance between plant growth and immunity, cold stress response	Zeng et al., 2022; Yuan et al., 2018
OG0004381	MYB			Several roles in development, biotic and abiotic response	Xiao et al., 2021; Wang et al., 2021
OG0009002	TCP	Species‐specific SOM	IBS cell area	Leaf and hypocotyl development	Aguilar‐Martinez & Sinha, 2013
OG0012289	GeBP			Defense response	García‐Cano et al., 2018

Abbreviations: 2°VB secondary vascular bundle; IBS, inner bundle sheath cells; M, mesophyll; M, mesophyll cell; Sb*, cell interface between bundle sheath cell and mesophyll cell per cross section area; SOM, self‐organizing map; VB, vascular bundle.

Finally, to identify features (genes displaced together with a particular phenotypic trait) that vary between species, the species‐specific SOM was built (Figure 6). Overall, the number of features displaced from C₃ to PK and PK to C₄ was similar; however, the nature of such displaced features differs between C₃/PK vs PK/C₄. Features displaced from C₃ to PK denoted a change in the timing of expression along development, with most exhibiting a delay in their expression pattern. Interestingly, traits such as leaf area, S_v*, IBS Area % and OBS cell area belonged to the delayed group of displaced features. This group is enriched in genes that denote high RNA metabolism. Features that may accelerate (early displacement) their expression pattern in PK compared to C₃ included VB diameter, OBS Area % and genes involved in catabolism, stress response and leaf development. Only three of a total of 15 phenotypic traits were conserved between C₃ and PK species, suggesting an important modification of leaf anatomy at this stage of evolution. While the number of features displaced from Proto‐Kranz to C₄ species showed similar trends to the displacement from C₃ to Proto‐Kranz, there was less difference between the number of these displacements. Most of the features displaced earlier in leaf development are enriched in genes involved in mRNA processing, among other functions. In contrast, features in the delay group are enriched in genes related to proteolysis, cell morphogenesis, and developmental cell growth. Interestingly, the transition from Proto‐Kranz to C₄ species suggests that most phenotypic traits are conserved, indicating that the leaf anatomy may not require significant adaptations for the C₄ pathway.

Addressing the displacement of leaf anatomical traits between species is another way of identifying potential regulators involved in diverse pathways. In particular, we were interested in getting insight into gene displacement accompanying changes in the BS area among Otachyriinae species. Our results show that the phenotypic trait BS cell area is conserved between C₃ and PK species but it is displaced in C₄ species (Figure 6). While BS cells reach their maximum area in S3 at the base of the leaf in C₃ and PK species, they do the same in S5, closer to the tip of the leaf, in the C₄. This delayed expression correlated with the displacement of 41 genes involved in cell wall modification and pyruvate transport, among others. Indeed, two genes in this group are known to be part of suberin biosynthesis: an ortholog to AtFAR1/4 (fatty acid reductase) a member of alcohol‐forming fatty acyl‐CoA reductases (Domergue et al., 2010) and an ortholog to AtABCG16, a class of ABCG half‐transporters that are required for the synthesis of an effective suberin barrier in roots and seed coats (Yadav et al., 2014). In addition, there are two transcription factors among the features displaced. One is a member of the class 1 TCP transcription factor family (AtTCP7) that in A. thaliana plays an important role during leaf and hypocotyl development (Aguilar‐Martinez & Sinha, 2013). The second one (AtBFP4) is a member of the GeBP family and is involved in the defense response in A. thaliana (García‐Cano et al., 2018). While the roles of these two candidate genes in leaf development remain unexplored in grasses, their association with suberin biosynthesis and defense response makes them promising candidates for further experimental investigations.

In conclusion, this is the first study showing how unsupervised machine learning can be used for the discovery of patterns of variations between phenotypic traits and genes associated with the development and evolution of divergent leaf anatomies. Moreover, this methodology is fast, easy to implement and permits visualizing hidden patterns in complex data matrices. This method can be easily applied to other model and non‐model species as well.

Deciphering gene network contributions to switches in phenotypic states, as observed in leaf development and evolution, involves identifying specific signaling gene factors. In this sense, the use of SOM allowed us to unlock genes related to leaf anatomical traits and to detect displacements in gene regulatory networks between Otachyriinae species with different leaf anatomies. Genes pinpointed here as potential drivers of phenotypic evolution are suitable candidates for further validation through functional studies in grasses. Finally, the work presented here increases our knowledge about the evolutionary mechanisms in photosynthesis and brings us closer to new comparative studies among other plant lineages.

AUTHOR CONTRIBUTIONS

SP, AJS, and RR defined the research question. SP, GS, and RR contributed to the study design and experiments; GS helped with the SOM model design. SP performed the experiments; SP, AJS, and RR contributed to the biological interpretation of the results; SP, GS, and RR wrote the manuscript; SP, GS, AJS, and RR revised the manuscript. RR is the corresponding author. All authors read and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

Supporting information

Data S1. Supporting Information

Table S1. Quality measures for SOMs.

Figure S1. Sampling the 5th leaf of four Otachyriinae subtribe species.

Figure S2. Sb* and Sv* parameters determination.

Figure S3. Developmental gradient in the 5th leaf of four Otachyriinae subtribe species.

Figure S4. Quantitative phenotypic traits.

Figure S5. SOM Interactive visualization graph.

Figure S6. SOM line visualization plot.

Figure S7. Species‐specific SOM.

Figure S8. Number of displacements.

Figure S9. Phenotypic traits displacement.

Dataset S1. Phenotypic traits and gene expression data used for SOM analysis.

Dataset S2. Membership list of neuron pairs (SOM 17x17) and REVIGO clustering of GO terms.

Dataset S3. REVIGO clustering of GO terms in reference neuron pairs.

Dataset S4. GO enriched terms (BP) in C3 to PK and PK to C4 displaced features.

Dataset S5. Genes displaced together with IBS cell area and GO (BP) enriched terms.

Prochetto, S. , Stegmayer, G. , Studer, A. J. , & Reinheimer, R. (2026). Systems analysis of leaf anatomy and transcriptome using unsupervised machine learning provides insight on photosynthesis development and evolution in non‐model grasses. Plant Direct, 10(3), e70022. 10.1002/pld3.70022

DATA AVAILABILITY STATEMENT

Data generated or analyzed in this study are included in this published article and its supplementary data published online at Dryad Repository (Prochetto, Santiago; Stegmayer, Georgina; Studer, Anthony J.; Reinheimer, Renata (2023), Prochetto et al. JEXBOT‐ Supplementary Dataset, Tables and Figures, Dryad, Dataset, https://doi.org/10.5061/dryad.p5hqbzktn). The Illumina RNA‐Seq reads and transcriptomes assembly underlying this article are available in NCBI Sequence Read Archive (Prochetto et al., 2023; PRJNA813546).