Novel tools for quantifying secondary growth

Abstract

Secondary growth occurs in dicotyledons and gymnosperms, and results in an increased girth of plant organs. It is driven primarily by the vascular cambium, which produces thousands of cells throughout the life of several plant species. For instance, even in the small herbaceous model plant Arabidopsis, manual quantification of this massive process is impractical. Here, we provide a comprehensive overview of current methods used to measure radial growth. We discuss the issues and problematics related to its quantification. We highlight recent advances and tools developed for automated cellular phenotyping and its future applications.

Arabidopsis, automated cellular phenotyping, machine learning, quantitative histology, secondary growth.

Introduction

Secondary growth, the radial thickening of plant organs, is a large-scale process: thousands of cells are produced by the vascular cambium throughout the life of most woody dicotyledonous plants and gymnosperms (Spicer and Groover, 2010; Ragni and Hardtke, 2014; Zhang et al., 2014). It occurs in the root, hypocotyl and stem of the herbaceous model species Arabidopsis thaliana (Arabidopsis). However, even in this relatively small model plant, soon after secondary growth begins, cell abundance is too large to perform manual quantifications effectively (Sankar et al., 2014).

Vascular anatomical disposition differs throughout the plant kingdom, depending on the species, the organ considered, and even the developmental stage. For instance, during secondary growth in Arabidopsis, the root and hypocotyl rearrange their vasculature from a diarch symmetry (with two opposite initial phloem and xylem poles) to a fully radial symmetry (with a ring of cambial cells that produces inward daughter cells, which will differentiate into xylem and outward daughter cells that will develop into phloem) (Esau, 1977; Dolan and Roberts, 1995; Chaffey et al., 2002; Ragni and Hardtke, 2014). In contrast, in the stem, secondary growth arises with the formation of the fascicular cambia in the 7–8 collateral bundles and later with the formation of an interfascicular cambium, which connects the bundles (Esau, 1977; Altamura et al., 2001; Ragni and Hardtke, 2014).

Another lateral meristem that contributes to the increase in girth of plant organs is the cork cambium or phellogen. It produces the phelloderm on the inner side and the cork or phellem tissues on the outer side. In many plant species, cork cells develop into suberized dead cells at maturity. Together with the phellogen, the phelloderm and phellem form the periderm (Esau, 1977). The periderm replaces the epidermis in stems, branches, and roots of most dicotyledons, and gymnosperms once the latter can no longer accommodate radial growth. The periderm acts as a protective barrier against biotic and abiotic stress (Pereira, 2011). Furthermore, the parenchymatic components of the phelloderm fulfill a function in storing starch (Esau, 1977).

The periderm can be studied in the root and hypocotyl of Arabidopsis (Dolan and Roberts, 1995; Chaffey et al., 2002). Working with the Arabidopsis hypocotyl offers several advantages because radial growth progression can be easily followed over time (as elongation and secondary growth are uncoupled) and the disposition of the vasculature is reminiscent of trees. Together these features make the Arabidopsis hypocotyl a good model to study secondary growth (Chaffey et al., 2002; Ragni and Hardtke, 2014). Briefly, hypocotyl secondary growth can be divided into two phases based on cell morphology and proliferation rate: an early phase in which xylem mainly comprises water-conducting cells and parenchyma, and a later phase of so-called xylem expansion in which xylem occupancy is increased and fibers differentiate (Chaffey et al., 2002; Sibout et al., 2008).

Many factors controlling vascular secondary growth have been identified (see the following reviews for more detail: Furuta et al., 2014; Zhang et al., 2014; Jouannet et al., 2015; De Rybel et al., 2016). However, not all the players are known, and the spatio-temporal regulation of radial growth is far from being understood. In this review, we will provide an overview of the issues posed by secondary growth quantification. Moreover, we will present approaches and tools that have the potential to advance the field.

Current approaches for the quantification of secondary growth

In tree species, overall secondary growth is traditionally quantified as stem diameter at reference internode positions, while more accurate analyses are achieved by measuring tissue widths, number of cells per tissue files, distances between tissues (i.e. distance from the outer bark to the pith), or a combination of these measurements (such as the ratio between the width of the wood and the stem radius) (Nieminen et al., 2008; Etchells et al., 2015; Miguel et al., 2016). Similarly, in the model plant Arabidopsis, overall secondary growth is quantified as diameter or the area of the different organs (the stem, the root, and the hypocotyl) (Altamura et al., 2001; Chaffey et al., 2002).

More specific parameters can be used to quantify Arabidopsis stem radial growth as the number of cells per vascular bundle, the tangential over radial ratio of vascular bundles, the lateral extension of the tissue produced by the interfascicular cambium, and the acropetal progression of interfascicular cambium initiation along the stem (Sehr et al., 2010; Agusti et al., 2011a, b; Etchells et al., 2012, 2013) (Fig. 1A). For the Arabidopsis hypocotyl and root, valid alternatives to diameter length are the xylem over total area ratio (xylem occupancy) (Fig. 1B) or the xylem over phloem area ratio (Sibout et al., 2008). More insights on xylem composition can be obtained by macerating woody samples to estimate the relative number of different cell types and their characteristics, such as shape and size (Franklin, 1945; Chaffey et al., 2002; Muñiz et al., 2008; Ragni et al., 2011) or by measuring the so-called ‘xylem 1’ (vessels and parenchyma) and ‘xylem 2’ (fibers and vessels) (Fig. 1C) (Chaffey et al., 2002; Liebsch et al., 2014).

Fig. 1.

Examples of secondary growth quantification. Cross-sections of plastic-embedded Arabidopsis: (A) Col-0 stem, 0.5 cm from the base of a 9-week-old plant; (B) Col-0 hypocotyl at 12 d after flowering. X/A, ratio between the xylem area and the total area; ICD, interfascicular cambium-derived tissue; * vascular bundle. (C) Vibratome section of Arabidopsis hypocotyl (Col-0, 15 d after flowering) stained with phluoroglucinol (in red) showing how ‘xylem 1’ (X1) and ‘xylem 2’ (X2) are measured. Scale bar=200 μm.

Challenges of secondary growth quantification

Many of the previously mentioned approaches only coarsely describe radial growth and do not capture its complexity at the morphological and temporal level. For instance, a reduction of hypocotyl area does not always reflect an overall reduction in cell proliferation. It could be due to small changes in cell sizes that cannot be easily detected by eye, or by an increased/decreased proliferation rate in one specific tissue. Along the same lines, both the presence of larger xylem vessels and more cell divisions could account for higher xylem occupancy in the hypocotyl radial section (Sankar et al., 2014; Lehmann and Hardtke, 2016). To be able to account for these growth patterns, it is necessary to track and quantify growth at a cellular level (Sankar et al., 2014).

However, manual quantification of secondary growth morphodynamics is impractical even in the tiny Arabidopsis plant, as there are >15 000 cell files in a mature hypocotyl. Moreover, the quantification of vascular morphodynamics is hampered not only by the scale of the process but also by the inaccessibility of certain tissues due to their deep location such as the xylem. Consequently, live imaging is challenging, and the majority of the measurements are achieved on cross-sections of embedded fixed samples. Thus, temporal resolution is also limited by sample preparation.

A further aspect to consider is that severe defects in radial growth often coincide with decreased plant viability. This renders the isolation of such plants challenging and the interpretation of those results even more difficult, as it is arduous to distinguish the direct contribution to secondary growth from pleiotropic effects. Thus, many studies rely on weak alleles or partially redundant contexts, in which plant fitness is not affected and phenotypes are mild. Therefore, quantification of secondary growth will benefit from novel tools for automated cellular phenotyping (Sankar et al., 2014; Lehmann and Hardtke, 2016).

Finally, it is important to emphasize that secondary growth studies in Arabidopsis should be normalized to the developmental stage rather than to absolute age, as flowering greatly influences secondary growth progression. For instance, it triggers xylem expansion and fiber formation in the root and in the hypocotyl (Sibout et al., 2008; Ragni et al., 2011; Ragni and Hardtke, 2014).

Quantitative histology

In the last few years, the increasing volume of genotyping data, generated using low-cost sequencing technology, has shifted attention from more efficient genotyping to more automated and precise phenotyping. However, while high-throughput plant phenotyping is well developed for laboratory and field experiments in model and crop plants, automated cellular phenotyping is still a novel field and it has only recently been applied to secondary growth (Sankar et al., 2014; Hall et al., 2016; Lehmann and Hardtke, 2016). Sankar et al. (2014) define ‘quantitative histology’ as an automated identification/quantification of cell types and cellular morphological descriptors in a tissue.

Several pipelines, such as RootScan, RootAnalyzer, PHIV-RootCell, and the method of Montengro-Johnson and colleagues exist for the characterization and quantification of primary root growth (Burton et al., 2012; Chopin et al., 2015; Lartaud et al., 2015; Montenegro-Johnson et al., 2015). These tools were developed to classify rice, wheat, and Arabidopsis roots, respectively. To ensure a high quality classification, these methods exploit a priori knowledge of root architecture (Burton et al., 2012; Chopin et al., 2015; Lartaud et al., 2015; Montenegro-Johnson et al., 2015). This renders their usage very specific and not easy to adapt to other species or organs. More recently, two methods for automating cell extraction, quantitative shape analysis and cell type classification (in any 2D tissue of interest), were developed independently (Sankar et al., 2014; Hall et al., 2016). Relying on generic machine learning (ML) methods, these protocols can be generalized and applied to tissues other than those used in the respective studies. For this reason, these methods can really form the basis of the so-called ‘quantitative histology’.

In more detail, both approaches rely on similar methodology, splitting the task into four steps: (i) image acquisition; (ii) image pre-processing; (iii) image segmentation and feature extraction; and (iv) cell type classification (Sankar et al., 2014; Hall et al., 2016) (Table 1; Fig. 2).

Table 1.

Key steps of ‘quantitative histology’ methods for secondary growth

Step	Description	Sankar et al. (2014)	Hall et al. (2016)
0	Sample preparation	Microtome sections Plastic-embedded tissue	Vibratome sections Calcofluor white staining
1	Image acquisition	Light microscopy Stitching/tiling	Laser confocal microscopy
2	Image pre-processing	Gamma contrast Gaussian blur	Gaussian blur
3	Segmentation and features extraction	Watershed algorithm Morphometric features	Watershed algorithm Morphometric features Fluorescence signal intensity
4	Cell type classification	Machine learning (SVM)	Machine learning (random Forest)

Fig. 2.

Example of the ‘quantitative histology’ approach. In (A–C) the same Arabidopsis hypocotyl section (Col-0 21 d after germination) is presented. (A) Row image, red box magnification showing details of the xylem, blue box magnification showing details of phloem. (B) Image after pre-processing and segmentation with a watershed algorithm; each color basin represents one cell. (C) Labeled image using a machine learning (ML) approach. Every color represents a cell type: yellow, xylem vessels (Xv); cyan, cambium (Cb); magenta, phloem elements (Phe); green, xylem parenchyma (Xp); blue, phloem parenchyma (Php); brown, periderm (Pe). (D) Rose diagram of the incline angles of the xylem vessels measured in (C). For instance, a value of 0 represents radial/anticlinal orientation, and a value of π/2 represents orthoradial/periclinal orientation. (E) Rose diagram of the incline angles of the cambial cells measured in (C). (F) Average of some features [eccentricity (Ec), area, perimeter (Per)] and cell numbers (n Cell) for each cell type measured in (C). Scale bar=100 μm.

Image acquisition is one of the most critical steps, and its importance has often been underestimated. The quality and the nature of the pictures acquired greatly influence the ease with which the other steps in the pipeline can be performed (i.e. images that are suited for segmentation required less pre-processing). A related point is the standardization of the image acquisition process among experiments; parameters will not change if the images are acquired in the same way and in the same conditions, allowing large-scale samples. Due to the fact that live imaging of thick organs, such as the hypocotyl during secondary growth, is still impossible with conventional microscopy, both methods rely on grayscale images of cross-sections of fixed samples. Thus, an additional step of sample preparation is required. The approach of Sankar et al. (2014) uses high-resolution images of plastic-embedded sections of fixed material, acquired using the tiling/stitching function of a microscope with a motorized stage (Fig. 2A). In contrast, Hall and colleagues use laser scanning confocal images of vibratome sections in which cell borders were outlined by fluorescent staining of the cell wall (Hall et al., 2016). Both strategies have advantages and disadvantages. Sample preparation for the first approach is easier at young developmental stages, and image resolution is higher, whereas the segmentation/pre-processing processes are slightly more difficult compared with confocal images (confocal images have less background and shadows). Moreover, only the procedure of Hall and colleagues allows the measurement of an additional fluorescent signal. A minor limitation of both protocols and still a general issue of secondary growth quantification is that radial growth is normally measured in cross-sections, and thus only in 2D (Lehmann and Hardtke, 2016).

After acquisition, the pre-processing transforms the images to improve the segmentation. This step is tightly linked to the imaging method and to the segmentation algorithm used. Both methods apply a series of filters to remove or reduce noise and reinforce the contrast. For instance, Sankar et al. (2014) use a Gaussian blur to filter out high-frequency noise, followed by a gamma adjustment to improve the general contrast of the image, while Hall et al. (2016) use only a blur for denoizing.

The segmentation is the process in which objects (in our case cells) are identified and extracted from the background and from each other. More precisely, a label is assigned to every pixel of the image, and pixels with the same label share certain characteristics (i.e. belong to the same cell). The two approaches rely on a common algorithm for the segmentation of grayscale images: the watershed algorithm (Fig. 2B) (Vincent and Soille, 1991; Yoo et al., 2002; Barbier de Reuille et al., 2005, 2015; Pound et al., 2012). Briefly, the watershed algorithm is based on the geographical concept of the watershed and catchment basin. In geography, a catchment basin is a region of a map in which water flows into the same lake or basin, while the watersheds are the limits at which water would enter different catchment basins. An image can be seen as a topographical surface where high pixel intensity corresponds to ‘high’ regions and low pixel intensity refers to ‘low’ regions; thus we can apply the same geographical definitions to this virtual map. For instance, if the cell wall is brighter than the cell content, we can identify cells as catchment basins for the segmentation process (Fig. 2B) (Vincent and Soille, 1991).

The accuracy of the segmentation is critical, as mis-segmented cells are likely to be wrongly classified. After the segmentation, cellular features/descriptors such as cell area, cell perimeter, position of the cell, and cell eccentricity are computed for every cell. The computation of the features is achieved using conceptually similar toolboxes (Pau et al., 2010) (http://www.diplib.org/main). Moreover, the incline angle—the angle formed by the major axis of the cell with the radius of the sample—is calculated in both protocols (Sankar et al., 2014; Hall et al., 2016). In addition, Hall et al. (2016) measured cell lumen area, and the cell wall area.

Cell type classification is achieved through an ML approach (Fig. 2C). The basic principle of ML is to teach computers to: (i) analyze existing data effectively; (ii) extract underlying similarity/differences; and (iii) generate a classifier/pattern to apply to new data (Bastanlar and Ozuysal, 2014; Ma et al., 2014; Libbrecht and Noble, 2015; Angermueller et al., 2016; Singh et al., 2016). The first step is the creation of a training set, a set of images that is used to learn the model, in which the cell types of interest are manually labeled. The second step is the choice of the features that better describe each class of cells. Then, different algorithms for supervised classification can be used to create the classifier. Sankar and colleagues used a Support Vectors Machine (SVM) algorithm, whereas Hall and colleagues reported to have better performances using a random Forest algorithm (please refer to the followingreviews for a comprehensive overview on ML algorithms: Bastanlar and Ozuysal, 2014; Ma et al., 2014; Libbrecht and Noble, 2015; Angermueller et al., 2016; Singh et al., 2016). Then, the classifier performances are tested against the so-called test set, a fraction of the training set, until they are satisfactory (several rounds of optimization may be necessary). Finally, the classifier is used to label cells (identify the different cell types) on a running set (a set of segmented images that was not manually labeled) (Fig. 2C). Sankar and colleagues also proposed an automated filter to correct mislabeled cells. Indeed, they reported that at later developmental stages, in some cases xylem identity was assigned to phloem parenchymatic cells. In such cases an automated quality check, based on manually created masks of the total area and of the xylem area, was implemented and used (Sankar et al., 2014).

In summary, the final output of the two approaches regarding identification of the cell type is rather comparable as both methods allow the quantification of similar features and cell types with accuracy >85% (Fig. 2F). However, the choice of the method should be dictated by the research focus/interests. For instance, we recommend the method of Hall and colleagues to study problems related to cell wall integrity, whereas we suggest the method of Sankar et al. for temporal analyses (developmental series with several time points).

Current and futures applications of ‘quantitative histology’

A proof of concept of ‘quantitative histology’ approaches is the detailed characterization of two Arabidopsis accessions that display differences in secondary growth progression: Ler and Col-0 (Ragni et al., 2011). The characterization of the morphodynamics between Ler and Col-0 revealed that overall secondary growth is more prominent in Col-0, whereas xylem occupancy is higher in Ler. This is not primarily due to cell size, although xylem cells in Ler are slightly bigger, but rather due to a decrease in phloem proliferation in Ler (Sankar et al., 2014). Another remarkable observation is that the spatio-temporal dynamics of the incline angles reflect the different phases of secondary growth of the Arabidopsis hypocotyl. For instance, at young stages (15 d), the inclines are uniformly distributed. At ~20 d, xylem cells are radial and cambial cells orthoradial, and the overall distribution starts to be bi-modal (Fig. 2D, E) (Sankar et al., 2014). In addition, this type of analysis pointed out some unexpected findings such as that the cambium produces more overall phloem than xylem, even though xylem occupancy is increasing during plant development, and that the enhanced xylem occupancy in Ler is not due to an increase of xylem cell numbers but mainly due to a decrease in phloem area.

Hall et al. (2016) validated their approach characterizing knotted1-like 1 (knat1)/brevipedicellus (bp) loss-of-function mutants. The knat1 mutant is suitable for a test as it exhibits reduced fiber cell number, combined with a decrease in xylem vessel area and altered cell wall deposition (Liebsch et al., 2014). In addition, they quantified xylan abundance across different cell types, coupling the intensity of a fluorescent signal (immunostaining of the xylan) with the cell type classification and the morphometric data. Other components of the cell walls (such as cellulose, lignin, and callose), which can be visualized by immunolabeling techniques, could be easily measured together with the cell morphological descriptors, paving the way for the analyses of the chemical composition of the cell wall with spatial resolution (Hall et al., 2016).

Another future application is to combine the automated cellular phenotyping with genome-wide association studies (GWAS). Strikingly, natural variation is still a largely untapped resource for the study of secondary growth. However, a large degree of variability in radial growth-related traits was observed in the hypocotyl of a small collection of Arabidopsis accessions, confirming the potential of this approach (Sibout et al., 2008; Ragni et al., 2011). Another aspect of secondary growth that can be further explored with more accurate phenotyping techniques is how secondary growth is modulated during changes of environmental conditioned and abiotic stresses.

So far, Sankar et al. (2014) and Hall et al. (2016) have tested their approaches only in the Arabidopsis hypocotyls. However, we expect to see the ‘quantitative histology’ approaches exploited in other plant species (tomato, poplar, etc.), as they are quite versatile and easy to adapt. For instance, we foresee minor tuning of the segmentation and machine learning parameters for the application to other organisms (as long as the images can be easily segmented). In addition, Hall et al. (2016) provide their method as a MATLAB package, with a graphical interface, and thus it does not require any coding by the user. Along the same lines, the ‘quantitative histology’ approach by Sankar et al. (2014) was recently implemented in the open source platform LithoGraphX [www.lithographx.org; a fork of MorphoGraphX (Barbier de Reuille et al., 2015)] to render it accessible to biologists. Other advantages of the implementation on this platform are: (i) the reduced computational time for the segmentation process; (ii) the possibility to use several types of images as input (laser confocal images, grayscale images, and color images) and several pre-processing tools; and (iii) the choice between the two ML algorithms (Barbier de Reuille and Ragni, 2017). Moreover, in LithoGraphX it is possible to perform the Hall et al. approach or other protocols, as LithoGraphX was initially developed for the analyses of confocal images and offers a variety of tools for measuring fluorescent signal intensity.

A possible future implementation is to add tissue-specific features to improve cell type classification of problematic tissues. For instance, phloem companion cells and sieve elements are difficult to distinguish from one another (Sankar et al., 2014). In fact the morphology of these cell types is nearly identical, which hampers the ML recognition process. Adding tissue-specific features such as the number of neighboring cells, cell wall thickness, or a particular stain should resolve this problem.

Conclusions and perspectives

In summary, it is fair to conclude that secondary growth characterization will benefit from precise quantification at the cellular level. ‘Quantitative histology’ paves the way towards the study of secondary growth with good spatio-temporal resolution: it facilitates the measurement of complex traits and mild phenotypes. We foresee that automated cellular phenotyping will boost natural variation studies and will soon be applied to other species.

To date, the rate-limiting step of ‘quantitative histology’ methods is sample preparation/image acquisition. This is especially true for secondary growth studies where sample preparation is laborious and not yet automatized. Any improvements in this direction will contribute to render the ‘quantitative histology’ approaches routine protocols. A major breakthrough will be to image live secondary growth progression. This will open the door to the study of secondary growth in 3D and 4D.

Acknowledgements

This work was supported by a DFG grant (RA-2590/1-1) and by the SystemsX.ch, the Swiss Initiative in Systems Biology. LR is indebted to the Baden-Württemberg Stiftung for the financial support of this research project by the Elite programme for Postdocs. We thank Dr Azahara Barra-Jiménez and Dr Kristine Hill for critical reading.

Author notes

* Correspondence: [email protected]

Editor: Simon Turner, University of Manchester