2.1. AFM-IR imaging of A. thaliana epidermal cell walls

The spatial resolution and signal quality of AFM-IR images depend greatly on sample thickness and roughness (Pereira et al., Reference Pereira, Flores-Borges, Bittencourt, Mayer, Kiyota, Araújo, Jansen, Freitas, Oliveira and Mazzafera2018). For example, in thick samples (>1 μm) of poor thermal conductivity, insufficient decay of thermal expansion with respect to scan rate can reduce spatial resolution and signal quality (Kurouski et al., Reference Kurouski, Dazzi, Zenobi and Centrone2020). Most AFM-IR studies have involved spun-cast polymer thin films rather than complex biological specimens. As a result, our study required special considerations for sample preparation.

Resin embedding is a commonly used technique to obtain thin sections with well-preserved cellular structures (Verhertbruggen et al., Reference Verhertbruggen, Walker, Guillon and Scheller2017). However, epoxy resin has strong IR absorption in the 1720–922 cm⁻¹ region, which would interfere with the IR signatures useful in our study (Coste et al., Reference Coste, Soliman, Bercu, Potiron, Lasri, Aguié-Béghin, Tetard, Chabbert and Molinari2021; González et al., Reference González, Cabanelas, Baselga and Theophanides2012; Pereira et al., Reference Pereira, Flores-Borges, Bittencourt, Mayer, Kiyota, Araújo, Jansen, Freitas, Oliveira and Mazzafera2018). Therefore, paraformaldehyde (PFA) tissue fixation and low melting temperature gelatin were used to embed apical portions of A. thaliana stems. A cryostat was then used to obtain cross sections from a frozen block of gelatin-embedded stem with controlled and uniform thickness (~ 10 μm) and surface roughness (<100 nm). This method resulted in sections that had intact epidermis, phloem and xylem cells, but softer tissues, like the pith and cortex, were less well preserved (Supplementary Figure S1A). AFM-IR imaging was conducted on the epidermal cell wall because the epidermis is composed of mostly primary cell walls (Cosgrove, Reference Cosgrove2018). Because sections were obtained perpendicular to the periclinal cell walls, the mechanical properties detected by AFM are unaffected by turgor pressure. PFA-fixed sections demonstrate minimal spectral difference from A. thaliana stem tissue that had undergone no sample preparation (Supplementary Figure S2). Therefore, the sample preparation method devised here was suitable for generating plant stem sections for AFM-IR with no detectable spectral contamination.

Fig. 1. AFM-IR imaging of an epidermal cell wall in a PFA-fixed A. thaliana stem sample. (a) Image of sample region from the AFM camera. AFM cantilever tip is outlined in yellow. (b,c) Contact resonance mode AFM height image and phase shift map of the large scan region (25 × 25 μm²) showing epidermal cells. Positive shifts in phase indicate increasing sample stiffness. Black box represents where AFM-IR measurements (30 × 30 spectral array) were collected. Scale bar = 5 μm. (d) Phase shift map of the small scan region (3 × 3 μm²) where AFM-IR measurements were collected. White dashed line is outlining a region of increased stiffness at the junction between two epidermal cells. (e,f) Zoomed-in images of two regions of varying phase shift along cell–cell junction, indicated in (d). Scale bar = 500 nm. (g) Height map was obtained of a 3 × 3 μm² region where AFM-IR data were collected.

Topographical (Figure 1a,b) and phase (Figure 1c) images obtained via contact resonance mode AFM imaging were used to locate an epidermal cell wall. The topographical map of a 25 $\times$ 25 μm² region shows the wall between two adjoining epidermal cells and corresponds to the AFM-IR image collected (Figure 1b). Phase images report the difference in AFM cantilever drive and response oscillation that is sensitive to the surface stiffness and adhesion of the surface to the AFM tip (Ye & Zhao, Reference Ye and Zhao2010). Generally, softer samples result in a negative phase shift and stiffer samples result in a positive phase shift (Dong & Yu, Reference Dong and Yu2003). Phase shift can be sensitive to cell wall composition and has been previously used to capture different cell wall ultrastructural features in young poplar stems (Farahi et al., Reference Farahi, Charrier, Tolbert, Lereu, Ragauskas, Davison and Passian2017). Similarly, in the A. thaliana sample from this study, the phase map (Figure 1c) showed striations along the wall, presumably with regions of softer and stiffer material. The spatial variation in sample surface and subsurface properties indicated by phase shift can be attributed to spatial variation in cell wall mechanics and/or adhesion (Bidhendi & Geitmann, Reference Bidhendi and Geitmann2019), which are thought to be determined by wall composition and organisation (Bidhendi et al., Reference Bidhendi, Altartouri, Gosselin and Geitmann2019; Bidhendi & Geitmann, Reference Bidhendi and Geitmann2019; Bou Daher et al., Reference Bou Daher, Chen, Bozorg, Clough, Jönsson and Braybrook2018; Cosgrove, Reference Cosgrove2018; Peaucelle et al., Reference Peaucelle, Braybrook, Le Guillou, Bron, Kuhlemeier and Höfte2011). For example, streaks in the phase shift maps indicate regions where the tip goes from an attractive (stiff) to a repulsive (soft) material (Figure 1d–f).

Chemical information on a shared epidermal cell wall was acquired with a 30 $\times$ 30 array (900 AFM-IR pixel points or spectra) over a 3 $\times$ 3 μm² region, collecting a matrix of spectra every 100 nm in the x- and y-dimension (black square in Figure 1b,c). Although the spatial resolution of our AFM-IR acquisition could not be determined, it nonetheless offered multimodal images to correlate cell wall chemical and mechanical properties. Within the scan region, images (256 $\times$ 256 pixel map over a 3 $\times$ 3 μm² region) were collected for phase shift (Figure 1d), height (Figure 1g), phase-locked loop (PLL) frequency (Supplementary Figure S1B) and deflection (Supplementary Figures S1C, E). Continuous adjustment in PLL frequency is necessary to maintain the resonance-enhanced condition throughout the scan. Hence, PLL monitors changes in contact resonance frequency which occurs due to changes in sample surface stiffness and adhesion (Stan & Solares, Reference Stan and Solares2014). Deflection is a direct measure of cantilever-surface interaction force and is used to produce a three-dimensional image of the sample surface (Vadillo-Rodríguez et al., Reference Vadillo-Rodríguez, Busscher, Norde, de Vries, Dijkstra, Stokroos and van der Mei2004). Based on Figure 1, the stiff portion of the wall, indicated by high positive phase shift, is outlined by a white dashed line (Figure 1d). This region appears slightly raised in the height map (Figure 1g), with more distinct three-dimensional structure in the deflection (Supplementary Figure S1C) and lateral deflection (Supplementary Figure S1D) maps. Based on the orientation of the sample, this region was determined to be the junction between the anticlinal cell walls of two epidermal cells. Figure 1e and f shows enlarged phase shift images of the cell–cell junction, suggesting a stiffer boundary between the cells and softer regions on either side. This stiff cell–cell junction likely represents the middle lamella, a pectin-rich region that acts as a ‘glue’ between shared cell walls (Daher & Braybrook, Reference Daher and Braybrook2015; Zamil & Geitmann, Reference Zamil and Geitmann2017). Previous work by Zamil et al. (Reference Zamil, Yi and Puri2014) using a micro-stretching technique also found that the middle lamella between epidermal cells is significantly stronger than the rest of the cell wall.

2.2. Determining the spatial variation in primary cell wall composition by AFM-IR

In our experiments, the available lasers for AFM-IR spectral acquisition were limited to a range of 922–1170, 1301–1410 and 1504–1720 cm⁻¹. All FTIR and AFM-IR spectra in this study were pre-processed applying spectral smoothing, baseline correction and normalisation approaches (Gautam et al., Reference Gautam, Vanga, Ariese and Umapathy2015) (see Supporting Information and Supplementary Figure S3). To ensure that spectra produced by AFM-IR are similar to those obtained by traditional FTIR, principal component analysis (PCA) was used to compare FTIR and AFM-IR spectra of carbohydrate standards obtained from plant primary cell wall. The results show that the FTIR and AFM-IR spectra of the carbohydrate standards are statistically indistinguishable (Supplementary Figure S2). Therefore, the AFM-IR spectra accurately reflect cell wall IR response and chemistry.

The IR spectra of the primary cell wall are complex because the different polymeric components each have intricate chemical functionality. FTIR spectra of cell wall standards were used to establish IR band assignments to specific primary cell wall components. Commercially available XG, RG-I, cellulose, xylan and polygalacturonic acid (PGA) samples extracted from various plant sources were used as representative examples of cellulose, pectin and hemicellulose. Lignin was not included as a cell wall standard since it is not abundant in the epidermal cell walls of A. thaliana apical stems (Zhu et al., Reference Zhu, Ganguly, Baskin, McClosky, Anderson, Foster, Meunier, Okamoto, Berg and Dixit2015). Further, imaging of lignin autofluorescence (Decou et al., Reference Decou, Serk, Ménard, Pesquet, de Lucas and Etchhells2017) demonstrated that there is essentially no lignin in the epidermis (Supplementary Figure S4). FTIR spectra of each standard are shown in Figure 2a, and major IR bands observed for each cell wall standard are detailed in Supplementary Table S1. To further facilitate interpretation of AFM-IR spectral data, FTIR spectra of A. thaliana primary cell wall reference fractions were obtained. Oxalate, 1 M KOH and 4 M KOH extracts were isolated from alcohol-insoluble residues generated from A. thaliana seedling by sequential extraction (Supplementary Figure S5). Previously, Peralta et al. performed sequential extractions followed by glycome profiling on the apical section of an A. thaliana stem (Pattathil et al., Reference Pattathil, Avci, Miller and Hahn2012; Peralta et al., Reference Peralta, Venkatachalam, Stone and Pattathil2017). It was found that the oxalate extract was enriched in RG-I and AG epitopes (Peralta et al., Reference Peralta, Venkatachalam, Stone and Pattathil2017). Similarly, in this study, the oxalate extract isolated from A. thaliana seedlings was found to consist mainly of Gal (39%), Ara (25%) and GalA (20%) residues (Supplementary Table S2). In addition to carbohydrates, this extract also contained a significant amount of protein (see Supplementary Table S3). Therefore, the oxalate extract, in this study, likely contains pectins, AGPs and other cell wall proteins. Peralta et al. (Reference Peralta, Venkatachalam, Stone and Pattathil2017) found that the subsequent carbonate and 1 M NaOH extracts were enriched in epitopes associated with xylans, HGA, RG-I and AG. In this study, the polysaccharides in the 1 M KOH extract were composed mainly of Xyl (35%), Ara (21%), Gal (20%), GalA (8%) and Glc (6%) residues (Supplementary Table S2). The 1 M KOH extract also contains a significant amount of protein (Supplementary Table S3). Therefore, the1 M KOH extract, in this study, likely contains xylans, pectins and cell wall proteins. Lastly, Peralta et al. (Reference Peralta, Venkatachalam, Stone and Pattathil2017) reported that a 4 M NaOH extract was enriched in epitopes associated with non-fucosylated XGs, fucosylated XGs and xylan. In this study, the polysaccharides in the 4 M KOH extract were composed mainly of GalA (25%), Xyl (24%), Ara (14%), Gal (13%) and Glc (10%) residues (see Supplementary Table S2). In addition, the 4 M KOH extract lacked significant amounts of protein (Supplementary Table S3). This dataset suggests that the 4 M KOH extract, in this study, likely consists of XGs, pectins and xylans.

Fig. 2. AFM-IR data of PFA-fixed A. thaliana stem epidermal cell wall show similar absorption patterns to FTIR spectra of cell wall standards. (a) Normalised FTIR spectra of cellulose, soy protein, PGA, RG-I, xylan and xyloglucan. FTIR absorption between 1,300–1,171 and 1,503–1,411 cm⁻¹ was removed for comparison with AFM-IR data. (b) A box plot of the AFM-IR spectra collected on the epidermal cell wall sample surface by binning 30 $\times$ 30 pixels into 15 $\times$ 15 pixels and by binning spectral wavenumbers 8 $\times$ . The red and blue lines represent the mean and median AFM-IR spectra, respectively. There was no AFM-IR absorption between 1503–1411 and 1300–1171 cm⁻¹.

Table 1 IR assignments of the average Arabidopsis AFM-IR spectrum compared to commercial standards.

The cellulose, xylan and XG standards exhibit strong IR absorption between ~975 and 1170 cm⁻¹ (Figure 2a). As shown in Table 1, these IR bands are related to: (a) sugar ring ether O-C-O, (b) primary and secondary alcohol C-O and (c) glycosidic ether O-C-O stretching commonly observed in the neutral residues and polysaccharide linkages of cell wall carbohydrates. The oxalate, 1 M KOH and 4 M KOH reference extracts, as well as the mean AFM-IR spectrum of the A. thaliana epidermal cell wall, also show strong IR absorption in the ~975–1170 cm⁻¹ region (Supplementary Figure S5 and Figure 2b), suggesting the presence of carbohydrates. The mean A. thaliana AFM-IR spectrum contains prominent bands with peak maxima at: (a) 1310 cm⁻¹ assigned to alcohol O-H bending and/or the amide III band; (b) 1380 cm⁻¹ assigned to alcohol O-H bending, aliphatic C-H bending and/or carboxylic acid symmetric stretching; (c) 1525 cm⁻¹ assigned to the amine II band and (d) 1670–1660 cm⁻¹ assigned to adsorbed water O-H bending, carboxylic acid antisymmetric stretching and/or the amine I band. Prominent IR bands at 1310, 1380, 1525 and 1657 cm⁻¹ are also observed for the soy protein standard (Figure 2a), indicating the presence of cell wall proteins in the A. thaliana epidermal cell wall spectrum. While the oxalate, 1 M KOH and 4 M KOH reference extracts also absorb at 1525 cm⁻¹ (Supplementary Figure S5), the relative absorption peak intensity at 1525–1075 cm⁻¹ for the 4 M KOH reference extract is much lower than for oxalate and 1 M KOH reference extracts. The lower intensity in this region suggests that the 4 M KOH reference extract contains significantly less protein. Pectin cell wall standards (i.e., PGA and RG-I) show strong IR bands at ~1407 and 1604 cm⁻¹, which are assigned to carboxylic acid symmetric and antisymmetric stretching (Figure 2a). Although IR absorption at ~1407 and 1604 cm⁻¹ is observed in the mean A. thaliana AFM-IR spectrum (Figure 2b) and in the oxalate, 1 M KOH and 4 M KOH reference extracts (Supplementary Figure S5), these bands overlap with the IR absorption of proteins, which makes resolving the presence of acidic sugars on pectin based on carboxylic acid functionality difficult. PCA analysis further specifies the extent of chemical similarity between cell wall reference fractions, cell wall standards and the A. thaliana epidermal cell wall regions (Supplementary Figure S6).

The variance of IR absorption intensity across measured wavenumbers is substantial (Figure 2b), suggesting that there is chemical complexity and heterogeneity as a function of AFM-IR pixel location. Single wavenumber maps (Figure 3) were generated to depict the spatial variation in IR absorption of specific functional groups associated with different cell wall standards (Table 1): neutral sugars in carbohydrates (C-O; 1075 cm⁻¹); amide II in proteins (C-N; 1525 cm⁻¹) and acidic sugar in carbohydrates, amide I in protein and/or adsorbed water (1660 cm⁻¹). While both 1075 and 1660 cm⁻¹ appear to have high absorption intensity across the majority of the AFM-IR scan region, there are clear differences in the level of absorption at these wavenumbers and their spatial distribution. For example, the area that shows the strongest absorption at 1075 cm⁻¹ shows weakest absorption at 1525 and 1660 cm⁻¹. Similarly, the region towards the bottom of the map that shows strong absorption at 1660 cm⁻¹ also shows weak absorption at 1075 and 1525 cm⁻¹. By contrast, in the top third of the map (region between 2.5 and 3.0 μm in the y-dimension), there is relatively equal spatial distribution of all three wavenumbers. Thus, single wavenumber maps can distinguish between the spatial distribution and concentration of spectrally resolved functional groups, such as C-O and C=O. However, these functional groups are found in multiple cell wall components. Therefore, single wavenumber maps are insufficient to fully capture the spatial and chemical complexity of cell wall components.

Fig. 3. IR wavenumber maps at (a) 1,075 cm⁻¹, (b) 1,525 cm⁻¹ and (c) 1,660 cm⁻¹. The black dashed outline represents the cell–cell junction highlighted in Figure 1d. a.u. = arbitrary units.

2.3. Deconvoluting AFM-IR spectra for mechanochemical correlations

To more accurately delineate local concentrations of specific cell wall functional groups and components, NMF spectral deconvolution was applied (Borodinov et al., Reference Borodinov, Bilkey, Foston, Ievlev, Belianinov, Jesse, Vasudevan, Kalinin and Ovchinnikova2019; Kulkarni et al., Reference Kulkarni, Dost, Bulut, Welle, Böcker, Boland and Svatoš2018; Labbe et al., Reference Labbe, Rials, Kelley, Cheng, Kim and Li2005; Lin et al., Reference Lin, Zhang, Wang, Li, Fan and Wang2018; Montcuquet et al., Reference Montcuquet, Hervé, Garcia, Dinten and Mars2010; Zhang et al., Reference Zhang, He, Yan, Gao, Ma and Li2018). NMF of the AFM-IR dataset (i.e., input matrix) requires a subjective, user-specified number of factors. The cophenetic correlation coefficient and residual sum of squares for a given solution represent the extent to which the residual data (i.e., the data from the input matrix that was not recreated by linear combinations of the factors) can be explained by best-fit linear combinations of solution factors. For example, Brunet et al. (Reference Brunet, Tamayo, Golub and Mesirov2004) proposed that the cophenetic correlation coefficients demonstrate the stability of NMF solutions or how accurately the NMF solution describes the dataset. Hutchins et al. (Reference Hutchins, Murphy, Singh and Graber2008) proposed that the proper NMF solution occurs when the residual sum of square, or the variation between the target matrix and the NMF solution, is minimised. However, these measures alone cannot solely inform which is the ‘best’ solution or number of factors. In this study, NMF solutions were generated with the number of factors ranging from two to six. Distinctive factor spectra were generated up to the three-factor solution, but following that point, increasing factor number resulted simply in factor splitting. Furthermore, the cophenetic correlation coefficient and residual sum of squares values were relatively low for the three-factor solution (Supplementary Figure S7). Therefore, the three-factor NMF solution was used to analyse the A. thaliana AFM-IR dataset.

Key IR absorption bands in the NMF spectra were analysed and used to attribute the origin of NMFs (Figure 4a and Table 2). Factor 1 spectrum exhibited strong absorption for sugar ring ether O-C-O and primary alcohol C-O stretching at IR bands between 975 and 1091 cm⁻¹, which are absorption signatures associated with cell wall carbohydrates (Figure 4a). Absorption in this region was also seen in the FTIR spectra of carbohydrate standards and reference extracts. Notably, similar to cellulose, XG and xylan, the Factor 1 spectrum did not show strong absorption in regions related to carboxylic acids (1407 and 1604 cm⁻¹). Factor 1 thus primarily represents a set of chemical groups that comprise neutral sugars on primary cell wall carbohydrate components, although there are likely spectral elements attributed to cellulose, XG and xylan in Factor 2. The Factor 2 spectrum contained IR absorption bands between (a) 1063 and 1170 cm⁻¹ with peak maxima at 1110 and 1165 cm⁻¹ that are attributed to glycosidic ether O-C-O and/or secondary alcohol C-O stretching and (b) 1301 and 1346 cm⁻¹ with a peak maximum at 1310 cm⁻¹ that is attributed to alcohol O-H bending and/or the amide III band. These two IR absorption band regions are found in the FTIR spectra of both carbohydrate and protein standards. Factor 2 thus represents specific moieties on primary cell wall carbohydrate components, mainly glycosidic linkages, and N-H groups of amides on proteins. The Factor 3 spectrum contained IR absorption bands between (a) 1324 and 1410 cm⁻¹ with peak maxima at 1380 and 1403 cm⁻¹ that are attributed to alcohol O-H bending, aliphatic C-H bending and/or carboxylic acid symmetric stretching, (b) 1504 and 1570 cm⁻¹ with a peak maximum at 1525 cm⁻¹ that is attributed to the amide II band and (c) 1596 and 1696 cm⁻¹ with peak maxima 1610 and 1660 cm⁻¹ that are attributed to carboxylic acid antisymmetric stretching, adsorbed water O–H bending and/or the amide I band. Factor 3 thus represents the acidic sugars on primary cell wall carbohydrate components, mainly carbonyls, as well as C-N and C=O groups of amides on proteins. The FTIR of RG-I and PGA suggest that the IR band for carboxylic acid antisymmetric stretching is centred closer to 1600 cm⁻¹. Accordingly, a major IR absorption feature of the spectra of Factor 3 located more towards 1660 cm⁻¹ is likely more characteristic of the presence of a protein amide I band or protein-adsorbed water. Factor 1 loading was mapped in Figure 4b. As expected for the plant cell wall, neutral sugars represented by Factor 1 are present throughout most of the AFM-IR scan region. The NMF loading maps for Factors 2 and 3 are shown in 6 × 6 pixel squares, in agreement with the cell wall being a composite material (Supplementary Figure S8) (Burton et al., Reference Burton, Gidley and Fincher2010).

Fig. 4. NMF factor spectra represent carbohydrate-rich and protein-like regions in the epidermal cell wall. (a) NMF factor spectra plot for the 3-factor NMF solution. Peaks in each NMF factor spectrum are labelled with vertical lines. There was no AFM-IR absorption between 1,503–1,411 cm⁻¹ and 1,300–1,171 cm⁻¹. (b) Distribution map of Factor 1 representing its concentration (absolute value) in each pixel, with stacked bar plots displaying the ratio of NMF scores of Factors 1, 2 and 3 binned across 6 $\times$ 6 pixels (one box of the grid). Black dashed outline represents the cell–cell junction highlighted in Figure 1d.

Table 2 IR assignments of NMF factors.

The left and top edges of the A. thaliana AFM-IR image are enhanced in the abundance of Factor 3 spectral features, whereas the bottom-right corner of the A. thaliana AFM-IR is enhanced in the abundance of Factor 2 spectral features. To better visualise these differences, we generated a map of the difference between Factor 1 score and the sum of Factors 2 and 3 scores (Figure 5a). Although an operation of this type with NMF loading values is ill-defined, it can be used to visualise which locations are enhanced in Factor 1 versus Factors 2 and 3 or vice versa. Positive values indicate locations where the IR spectra are enhanced in Factor 1, while negative values indicate locations with IR spectra are enhanced in Factors 2 and 3. In this map, Factor 1 dominates an oval-shaped region along the cell–cell junction (cyan and blue pixels), suggesting a high relative abundance of neutral sugars in this region. This map suggests that cellulose, XG and xylan are present at higher concentrations than proteins and pectins within this region. By contrast, proteins and pectins represented by Factors 2 and 3 together appear to dominate the bottom third of the AFM-IR image (yellow and red pixels). Plants contain various cell wall proteins that make up 5–10% of the cell wall (Albenne et al., Reference Albenne, Canut and Jamet2013; Amos & Mohnen, Reference Amos and Mohnen2019; Daher & Braybrook, Reference Daher and Braybrook2015; Keller, Reference Keller1993). Localised deposition of cell wall proteins was previously demonstrated by Smallwood et al. (Reference Smallwood, Martin and Knox1995), who performed immunohistochemical staining of extensin and showed it localises as puncta in the root epidermal cell wall. Without the need for immunostaining, our results further support that cell wall proteins may be preferentially deposited in certain regions of the cell wall. The remaining light-green coloured pixels in the map depict an equal mixture of Factor 1 and Factors 2 + 3, indicating that in some regions of the cell wall the concentration of cell wall components is more balanced. These results show that NMF is a useful statistical tool to deconvolute AFM-IR spectral datasets to identify the major chemical components of cell walls and to distinguish their spatial organisation and concentration.

Fig. 5. Higher relative presence of NMF Factor 1 positively correlates with cell wall stiffness. (a) NMF distribution difference between Factor 1 score and sum of Factors 2 and 3 scores. (b). Correlation between phase shift and the NMF distribution difference between Factor 1 score and sum of Factors 2 and 3 scores as seen in Figure 4a for each pixel. Black dashed outline represents cell–cell junction highlighted in Figure 1d.

2.4. Correlating cell wall composition and stiffness

An advantage of AFM-IR is its use in co-registering chemical and nanomechanical information. Cross-correlation coefficients can be used to evaluate the relationship between local mechanical properties, measured by cantilever phase shift, and chemical fingerprints (Borodinov et al., Reference Borodinov, Bilkey, Foston, Ievlev, Belianinov, Jesse, Vasudevan, Kalinin and Ovchinnikova2019). In this study, a positive cross-correlation value at a given pixel could result from either an increase in both the relative Factor 1 score and phase shift or an increase in relative combined Factors 2 and 3 scores and decrease in phase shift. Conversely, a negative cross-correlation value at a given pixel could result from either an increase in relative Factor 1 score but decrease in phase shift or an increase in both the relative combined Factors 2 and 3 scores and phase shift. The former result would suggest that cell wall components described by Factor 1 correlate with cell wall stiffening while the cell wall components described by Factors 2 and 3 correlate with cell wall softening, while the latter result would suggest the reverse. The magnitude of the cross-correlation value is indicative of the strength of correlation.

The epidermal cell–cell junction displayed a distinguishable topological feature (Figure 1g) as well as appeared stiffer than the rest of the scanned region, as indicated by positive phase shift (Figure 1d). Cross-correlation analysis showed that higher relative abundance of neutral sugars represented by Factor 1 correlates with cell wall stiffening, specifically along the cell–cell junction (blue pixels; Figure 5b). It is possible that around the cell–cell junction, a higher concentration of crystalline cellulose is driving wall stiffening (Rongpipi et al., Reference Rongpipi, Ye, Gomez and Gomez2019). Combined Factors 2 and 3 scores correlate with cell wall stiffening in a small region along the cell–cell junction (red and yellow pixels; Figure 5b). NMF factor score and phase shift in regions adjacent to the cell–cell junction are not strongly correlated, suggesting that softening of the cell is not solely due to the concentration of a particular cell wall component or, at least, the loading of a Factor 1 or Factors 2 + 3.