Plant material and cultivation

Plants were grown in commercial substrate (Klasmann-Deilmann GmbH, Germany) for 10 days in the greenhouse at 23°C daytime and 20°C night with a photoperiod of 16 hours. The primary leaves were then detached and transferred onto phyto agar plates containing 0.34 mM benzimidazol. H. vulgare cv. Ingrid wild type (WT) was used as a genotype susceptible to powdery mildew. The corresponding near-isogenic line Ingrid M.C. 20, containing dysfunction in mildew locus o 3 (mlo3) [26] was used to assess non race-specific papilla-based resistance. H. vulgare cv. Pallas 01, with resistant mildew locus a 1 and a 12 (Mla1; Mla12) was used to analyze a hypersensitive response [27, 28].

Pathogen and inoculation

The barley mlo3 and Mla1 avirulent Bgh isolate K1 was used to analyze resistance reactions 0–120 hours after inoculation (hai) [29, 30]. Bgh K1 was maintained on cv. Tocada (KWS, Einbeck, Germany) in a controlled environment. Twenty-four hours before inoculation the conidia of heavily infested plants were shaken-off and discarded in order to assure homogenous and vital conidia for inoculation. Detached leaves on phyto agar were inoculated with a density of = 329 (± 107) conidia/cm² from young powdery mildew pustules (7–10 days after inoculation (dai) with B. graminis f. sp. hordei). The agar plates were sealed and incubated in a climate chamber at 19°C, 1100 m⁻² cd illuminance and a photoperiod of 16 h per day. Development of Bgh and processes during resistance response of barley were histologically assessed, using Coomassie Brilliant Blue R-250 (CBB) and 3,3’-diaminobenzidine (DAB) staining [31, 32].

Total RNA extraction, reverse transcription-polymerase chain reaction and real time qPCR

For each time point and treatment, RNA was isolated and purified in three independent experiments from five randomly pooled LN₂ homogenized barley leaves (~1 g powdered material) using NucleoSpin RNA Plant kit and NucleoSpin RNA Clean-up kit (Macherey- Nagel, Düren, Germany), according to the manufacturers recommendations. Total RNA was determined by qPCR using GoTag G2 Hot Start Colorless Master Mix (Promega, Madison, USA). Therefore, the housekeeping genes UBIQUITIN and Actin were amplified, including genomic DNA as control sample. Subsequently, the qPCR product and RNA samples were stained with ethidium bromide in 2% agarose gel to detect contaminations with genomic DNA. Samples with DNA contamination were digested with DNase using TURBO DNA-free kit (Life Technologies, Carlsbad, USA) followed by the previous control step. The amount and quality of RNAs were photometrically analyzed using NanoDrop 2000 UV-VIS spectrophotometer (Thermo Scientific, Wilmington, USA).

For cDNA reverse synthesis 1 μg RNA was used. High-Capacity cDNA Reverse Transcription kit with RNase inhibitor (Applied Biosystems, Foster City, USA) was used for reverse transcription polymerase chain reaction following the manufacturer’s instructions. The cDNAs were checked by ethidium bromide staining in 2% agarose gel. Before performing real time qPCR, cDNA was diluted 1:50 with nuclease free water. PCR program was used for the Fast SYBR Green Master Mix (Applied Biosystems, Foster City, USA) with adjusted annealing temperature to the primers used. The amplified genes of interest were quantified in a relative way to the appropriate mock samples and the fold change in gene expression was calculated via the 2^−ΔΔCt -method by Livak and Schmittgen [33], using UBIQUITIN cDNA amplification for normalization. Used DNA oligonucleotide primers are shown in supplementary S1 Table. Amplified real-time PCR products were proved on 2% agarose gel after melting curve analysis at 55 − 95°C. Significant differences within the gene expression profile of the genes are analyzed by a Tukey-HSD Test with alpha = 0.5.

Hyperspectral image acquisition, reflectance extraction and histological observation

Hyperspectral images were acquired with a hyperspectral line scanner (ImSpector PFD V10E, Spectral Imaging Ltd., Oulu, Finland) within a spectral range of 400 to 1000 nm and a spectral resolution of 2.73 nm, installed on a microscope [34]. Hyperspectral measurements were performed without ambient light. Samples were illuminated with two linear light emitters (Dual line Lightlines, Schott, Mainz, Germany) in a vertical orientation of 30° and a distance of 20 cm to the sample besides the fore optic. As a light source, a 150 watt halogen tungsten lamp connected to the line lights via a non-absorbing fiber was used (DCR Light Source EKE, Polytec, Waldbronn, Germany). Hyperspectral Spectral binning and spatial binning were set to 1. Frame rate and exposure time were adjusted to the object. A magnification of 7.3x (~7.5 μm per pixel spatial resolution) was used. Hyperspectral data cubes were assessed from four healthy and eight with powdery mildew inoculated leaves per genotype, every 3 hours from 0 hai until 48 hai, followed by daily measurements until 120 hai. To receive the relative reflectance a white reference bar (SphereOptics GmbH, Uhldingen-Mühlhofen, Germany) with a 100% reflection was recorded (W), followed by a dark current image (B₁) to inherent the irradiance of the sensor. Subsequently, the leaf sample (I₀) and a corresponding dark current image (B₂) was recorded. Calculation of relative reflectance was according to the formula (I = (I₀-B₂)/(W-B₁)), using the software ENVI 5.1 + IDL 8.3 (ITT Visual Information Solutions, Boulder, USA). Following, spectral signals were smoothed by employing the Savitzky-Golay filter [35] and cutting reflectance values below 420 nm and above 830 nm because of high measurement noise. Parameters for the Savitzky-Golay filter were 25 centered supporting points and a third degree polynomial.

Spectral signatures of pixels from healthy and diseased regions were extracted manually. Therefore, a region of interest of about ~1,000,000 pixels was extracted for every image of all screened leaves. The arithmetic average spectrum was calculated for the regions of interest as the basis for further analysis. Differences in hyperspectral reflectance among genotypes and over time were calculated according to Carter and Knapp [36]. Data mining approaches inspired by Kuska et al. [34], and Kuska et al. [37], were used to compute binary maps for Bgh disease detection on WT and mlo3 barley as well as HRs on Mla1. Simplex Volume Maximization (SiVM) was applied to determine extreme signatures from hyperspectral images and all signatures were represented as a linear combination of these selected extremes. Thereby, these extremes are real spectral signatures, that are interpretable and uncover the variation existing in the data. In addition, the new representation enables expression of specific signatures as level of transition between healthy and diseased leaves [11, 34, 38] and the algorithm scales well with data dimensionality and size enabling statistical data mining on a massive scale e.g. in high-throughput hyperspectral imaging of plants [39]. In these experiments, the number of extremes computed by SiVM were set to k = 25. Differences in probability based on the simplex distributions of healthy and inoculated samples were then used for highlighting potentially affected regions.

Linear Discriminant Analysis for the visualization of synergetic effects

Linear Discriminant Analysis (LDA) can reduce the dimensionality of a dataset retaining the separability of the classes within the dataset. Visualizing the low-dimensional representation reveals the suitability of the features to separate the classes. This allows classification of an unknown sample as a member of a particular group. As an example, spectral investigation with different fruits applied successfully the LDA to distinguish fruits and vegetables based on the spectral properties of the parenchymal cell walls [40]. In addition due to its low model complexity, LDA is common in bio-statistical approaches with many generalizations and regularizations e.g. in microarray analysis [41].

LDA included in Matlab 2013a (The MathWorks Inc., Massachusetts, USA), was applied to three different views on the plant status: (I) spectral reflectance, (II) gene activity and (III) the combination of both. For the calculation of the LDA transformations to two dimensions, each spectral observation is represented by five evenly distributed bands and each gene activity measurement is represented by three technical repetitions. The combined feature set includes all features using feature stacking. Direction of the vectors was calculated by changed center of gravity during the experimental period.

RELIEF algorithm for feature selection

The selection of relevant features is an important preprocessing step for the reduction of evaluation time and improvement of result quality in hyperspectral data analysis. For the assessment of feature relevance, multiple methods were developed [42]. Within this study the regressional RELIEF-F (Relief) algorithm for feature selection included in Matlab 2013a (The MathWorks Inc., Massachusetts, USA) was used to gather additional information about the relationship between single spectral bands and plant-physiological properties (Fig 1). The Relief algorithm as a filter algorithm, determines the relevance of features independently from a specific prediction algorithm. The algorithm is based on the idea that the values of relevant features are not arbitrarily distributed in the feature space but that they are grouped by corresponding classes [43]. By analyzing the local neighborhood for all data points of a data set, the local characteristics can be transferred to the suitability of a single feature for a specific prediction tasks [44]. In the classification case (Fig 1), the difference Δ in a feature value f_j of a data point X_i to the next data point from its class H_i (hit) and from a different class M_i (miss) is transferred to the relevance R_i(f_j) = Δf_j(X_i,M_i)−Δf_j(X_i,H_i). Averaging the relevance across all n data points X_i, i = 1…n gives a global estimation of feature relevance based on the local characteristics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Linking hyperspectral imaging data with gene expression profiles.

Healthy and B. graminis f.sp. hordei inoculated primary leaves were used for hyperspectral imaging and RNA extraction. After hyperspectral image acquisition and normalization, the Savitzky-Golay-filter was applied to reduce measurement noise. Reflectance values were manually extracted from regions of interest and used to identify relevant time points in the barley-Bgh interactions by applying the Relief algorithm. Spectral reflectance of identified relevant time points were used to calculate spectral differences between healthy and Bgh inoculated barley. In parallel, RNA was used for reverse transcriptase to synthesize cDNA and to analyze gene expressions of selected genes using real time PCR. Linear Discriminant Analysis (LDA) was applied on the hyperspectral data, gene expression profiles and a combination of both data sets to visualize the suitability of the features for the separation of different barley-Bgh interactions. To determine relevance as indicator for interpretable relations between hyperspectral signatures and gene expression profiles the Relief algorithm was applied (blue = high relevant wavelength band).

https://doi.org/10.1371/journal.pone.0213291.g001

In this work, the Relief algorithm was used to determine the relevance as an indicator for interpretable relations within the data. It was applied to investigate the relevance of single wavebands of each genotype after inoculation. The significance of these differences were tested by applying the Welch's t-test (Supplementary S1 Fig). In addition, it was also possible to evaluate differences in the relevance of gene expression activities for the hyperspectral reflectance characteristics. This method has been used to select informative genes for cancer classification using microarray gene expression data and for the development of spectral indices for plant disease detection [45, 46].

Relevance of spectral wavelengths for discrimination of non-inoculated and inoculated barley

To discriminate between healthy and Bgh inoculated leaves, relevant wavebands were determined using the Relief algorithm (Fig 2A, 2C and 2E). Therefore, the hyperspectral reflectance of healthy and Bgh inoculated leaves were used as one data set in the Relief algorithm. The relevance of a waveband, to distinguish between healthy and Bgh inoculated, is indicated from blue (low relevance) to yellow (high relevance).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Identifying relevant time points during different barley-powdery mildew interactions.

a,c,e, Relevant wavelength bands during susceptible WT as well as resistant mlo3 and Mla1 barley-B. graminis f.sp. hordei (Bgh) interactions 0–120 hours after inoculation (hai). Dashed lines indicate changed measurement frequency in the experimental period. b,d,f, Corresponding RGB images were determined to analyze the phenotype 0, 24, 48 and 72 hai. Histological observations using Coomassie Brilliant Blue R-250 (CBB) and 3,3’- diaminobenzidine (DAB) were used to assess the development of Bgh and resistance responses of barley. Arrows indicate formed papilla in mlo3 leaves (n = 8 replicates. Scale bars correspond to 100 μm).

https://doi.org/10.1371/journal.pone.0213291.g002

Hyperspectral reflectance from 0 to 120 hai were analyzed to identify key time points of changes in spectral reflectance for the different interactions. The compatible interaction (WT) reveals relevant wavelength from 520–660 nm and around 690 nm, at 6 hai and 48 hai (Fig 2A). Waveband range of 400–520 nm shows high relevance 72 hai and 120 hai. The near infrared (NIR, 700–830 nm) range indicates higher relevance 36–48 hai and 96–120 hai, especially around 720 nm 120 hai. Secondary mycelia have grown at 72 hai on the WT, visualized by CBB staining (Fig 2B) and no significant H₂O₂ generation could be observed 72 hai. Visible powdery mildew symptoms appeared 120 hai (not shown).

Bgh inoculated mlo3 leaves showed high spectral relevance in wavelengths from 400–720 nm 0–48 hai (Fig 2C). Wavebands with the highest relevance to discriminate between healthy and inoculated mlo3 leaves were around 570 nm 36 hai and 48 hai. Later time points show similar relevance over the entire spectrum. The leaves did not show any symptoms. Microscopic observations at 72 hai by CBB and DAB staining indicated fully developed papilla, which stopped Bgh penetration and fungal development (Fig 2D). Wavelengths with high relevance to determine inoculated and to distinguish healthy Mla1 leaves, were indicated by the Relief algorithm at 400–600 nm and around 730 nm 72–120 hai (Fig 2E). Wavelengths in the blue range provide relevance 12 and 24 hai. H₂O₂ generations were microscopically confirmed 72 hai using DAB staining, which indicates HRs against Bgh (Fig 2F). After Bgh penetration, pathogen development was stopped by HR at Mla1 genotypes (Fig 2F, CBB).

Overall, the Relief algorithm identified relevant changes in the spectral reflectance during early plant-pathogen interaction for all three barley-powdery mildew interactions specifically 12, 24, 48, 72 and 120 hai. The significance of spectral changes at these time points was analyzed using the Welch’s t-test (Supplementary S1 Fig). These specific time points were further used to calculate differences in the hyperspectral reflectance between healthy and Bgh inoculated leaves to assess subtle changes during barley-Bgh interactions.

Early hyperspectral reflectance differences between susceptible, mlo and Mla resistant barley during interaction with Bgh

To identify differences in the spectral pattern during barley-Bgh interactions, the difference of hyperspectral reflectance intensity between healthy and Bgh inoculated leaves were calculated (Fig 3). Spectral differences were calculated for 0.5, 12, 24, 48 and 72 hai. At these time points symptoms were not macroscopically visible (Fig 3). First symptoms occurred 120 hai for the naked eye (not shown). For this reason, this time point was used as a macroscopically positive control but was no longer considered for further analysis. Instead, spectral signatures at 0.5 hai were used to assess influence of Bgh inoculation on the spectral reflectance. Calculations of difference spectra allowed an explicit analysis of significant wavelength ranges between healthy and inoculated genotypes (Fig 3). All minus values correspond to a higher reflectance intensity of healthy leaves and all plus values to a higher reflectance intensity of Bgh inoculated leaves. Differences between healthy and inoculated WT leaves were marginal around ~0.0024 [Δ%/100] mainly in the NIR range 0.5 hai (Fig 3A). The reflectance intensity of healthy WT leaves increased 400–670 nm and around 710 nm 12–48 hai. During these points in time, reflectance in the NIR range was slightly higher in WT inoculated leaves. Conidia already penetrated epidermal cells and haustoria were developed. The spectral pattern changed by an increased reflectance at 490–670 nm and around 700 nm and a decreased reflectance in the NIR 72 hai, investigated by the calculated difference signature (Fig 3A). At this timepoint, Bgh is in the post penetration phase including first full developed haustoria and secondary mycelia. Computed disease maps indicate powdery mildew infested spots 72 hai.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Assess spectral reflectance signatures by difference calculation.

a,b,c Spectral differences of healthy and inoculated with B. graminis f.sp. hordei (Bgh), susceptible WT, resistant mlo3 and Mla1, 0.5, 12, 24, 48 and 72 hai. Positive values in the difference plot demonstrate higher reflectance intensity of healthy leaves, negative values higher reflectance intensity of inoculated leaves. In addition, schemes of the interaction types development are illustrated. On pseudo RGB image from HSI, no symptoms are visible. Based on the reflectance spectrum, computed disease maps indicate Bgh infested pixels in white 72 hai. HR maps indicated pixels, which undergoes a hypersensitive response also in white 72 hai. (n = 8 biological replicates).

https://doi.org/10.1371/journal.pone.0213291.g003

Spectral differences between healthy and inoculated mlo3 leaves exhibit a different pattern (Fig 3B). Reflectance intensities were higher on healthy leaves over the entire spectrum 0 to 48 hai. During this time span, effective papillae developed in mlo3 inoculated leaves, but only marginal changes in the spectrum could be observed. Highly significant spectral changes were shown 400–720 nm 72 hai (Fig 3B and Supplementary S1B Fig). The disease map of inoculated mlo3 shows only black pixels, indicating healthy tissue.

Bgh inoculated Mla1 show higher reflectance intensities at 500 to 850 nm from 0.5 until 48 hai (Fig 3C). Early differences in reflectance were found in the NIR especially around 710 nm and 780 nm with a difference up to 0.027 [Δ%/100] 12 hai. During this time point of interaction, Bgh has developed an appressorium and starts to penetrate the epidermal cells and subsequently on the Mla1 genotype a HR is induced. During the HR process, spectral difference pattern changed 72 hai (Fig 3D). These spectral changes were highly significant from 500–650 nm and 690–830 nm (Supplementary S1C Fig). HR spots were not visible for the naked eye until 72 hai. Nevertheless, the applied HR map indicated pixels undergoing a HR in white.

Gene expression profiles during barley-powdery mildew interactions

In order to elucidate transcriptional activity of near isogenic barley lines during Bgh interaction, seven genes related to senescence and cell metabolism, cell signaling and resistance response were analyzed 0.5, 12, 24, 48 and 72 hai. Bgh inoculated WT showed down-regulated senescence corresponding genes indicated by a negative change of HvDRF and HvRuBisCO (Fig 4A). HvDRF was down-regulated from 24 until 72 hai. In contrast, HvRuBisCO was down-regulated already 12 until 72 hai without significant differences compared to healthy leaves 48 hai. Resistance responses by WT leaves against Bgh were indicated by up-regulated PR genes 12 and 72 hai. Differences in transcriptional activity between HvPR2, HvPR3 and HvPR5 during 24 and 48 hai were present. Interestingly, HvGlb1 showed higher transcriptional activity 12 and 48 hai in inoculated WT leaves according to Bgh penetration time points. Down regulation of HvJIP23 from 24 hai suggest that jasmonic acid is not an early key factor in this interaction system.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Gene expression profiles of seven genes affected during different barley-powdery mildew interactions.

a,b,c, Gene expression profiles of inoculated with B. graminis f.sp. hordei (Bgh), susceptible WT, resistant mlo3 and Mla1, normalized to their appropriated mock samples 0.5, 12, 24, 48 and 72 hai. (n = 3 replicates; error bars indicate standard deviation; asterisks show significant differences within the gene expression profile per gene calculated with Tukey-HSD, α = 0.05).

https://doi.org/10.1371/journal.pone.0213291.g004

Gene expression-profile for mlo3 inoculated leaves exhibit a different pattern compared to the other genotypes and treatments (Fig 4B). HvPR3 codes a chitinase 2 and was already upregulated 0.5 hai. All tested genes were then down-regulated 12 hai, except HvPR2 which had a higher transcriptional activity until 48 hai. HvDRF, HvRuBisCO and PR genes had a higher transcriptional activity in inoculated leaves compared to healthy mlo3 leaves 48 hai.

Gene expression-profile of Mla1 inoculated leaves refer to an initiate cellular degradation approved by down regulated gene expression of mostly all genes 12 hai (Fig 4C). HvGlb1 and HvJIP23 were the only genes in this system, which were significantly up regulated 48 hai.

Interrelations between hyperspectral signatures and gene expression profiles

To discriminate early Bgh pathogenesis on WT leaves and resistance responses on mlo3 and Mla1 leaves, LDA was applied on hyperspectral signatures, gene expression profiles and combined data sets (Fig 5). The LDA classification by hyperspectral signatures revealed an improved discrimination of the barley-Bgh interaction types with ongoing pathogenesis (Fig 5A). In contrast, LDA classification of the gene expression profiles of Bgh inoculated barley genotypes carried out a discrimination during the early time points (Fig 5B). Applying the LDA on the combined data sets of hyperspectral signatures and gene expression profiles revealed an efficient discrimination between the susceptible WT, mlo3 and Mla1 based resistant barley (Fig 5C). This shows significant synergistic effects between the hyperspectral signal and the corresponding gene activities carving out the subtle differences between the early plant adaptions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Feature detection by Linear Discriminant Analysis (LDA).

a,b The suitability of features to differentiate Bgh inoculated susceptible wild type (WT) barley, papilla resistant (mlo3) and hypersensitive resistant (Mla1) barley were tested using a LDA on hyperspectral signatures, gene expression profiles and the combined data sets 0–72 hai. c, Combining both data sets increased the separation between the different interaction types WT, mlo3 and Mla1. (Arrows indicate the changed centre of gravity per barley-Bgh interaction over the experimental period).

https://doi.org/10.1371/journal.pone.0213291.g005

Relevant wavelength bands for the individual gene expression profiles

To analyze the coherence between spectral reflectance and gene expression profiles during barley-Bgh interactions, relevant wavelength bands and reflectance intensities for each gene were computed using the Relief algorithm. Wavebands with high relevance are indicated in dark red, low relevance is indicated by beige color (Fig 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Relief algorithm to assess differences in the relevance of individual genes per wavelength band.

a,b,c High relevant wavelength bands for specific gene expression profiles of barley genotypes inoculated with B. graminis f.sp. hordei (Bgh) are indicated in dark red and low relevant wavelength bands in beige.

https://doi.org/10.1371/journal.pone.0213291.g006

Relevant wavebands for HvDRF expression profile of Bgh inoculated WT are around 730 nm (Fig 6A). The NIR range is relevant for the expression profile of HvRuBisCO, HvJIP23 and HvPR5. Wavebands at 400–530 nm and around 680 nm are relevant for HvPR5 expression during WT-Bgh interaction, respectively.

Expression profile of HvRuBisCO of Bgh inoculated mlo3 had varying relevant wavebands in the VIS-NIR range with the highest relevance around 730 nm (Fig 6B). High relevance around 730 nm and 760–830 nm are shown as well for HvJIP23. HvPR3 expression profile was related to reflectance at 400–530 nm, around 690 nm and 730 nm.

Relevant wavebands for gene expression profiles of Bgh inoculated Mla1 leaves were mainly computed from 660–680 nm for almost all genes. Thereby, HvDRF profile shows no relevant wavebands in the VIS-NIR range. Wavebands from 520–650 nm, 660–680 nm and 700–750 nm show high relevance for the expression profile of HvPR5.

Bgh influence early spectral characteristics of barley

Previous studies on hyperspectral imaging of barley-Bgh interactions elucidated spectral characteristics of powdery mildew pathogenesis, as well as for Mla1 and mlo3 based resistance between 0 until 14 dai [34, 38]. Important time points varied depending on the interaction. Investigations by Kuska et al. [37] indicated key moments in the plant-pathogen interaction to distinguish barley-Bgh interactions by hyperspectral imaging and histologic analysis 2–3 dai. In the present study the relevance of spectral changes during barley-Bgh interactions were revealed for the first time. Hyperspectral imaging of early time points (0–48 hai) revealed changes in the metabolism of inoculated WT and mlo3 near-isogenic lines. During this interaction time, senescence was decelerated in inoculated leaves, indicated by stable reflectance intensity in the green range.

Barley leaves with a dysfunctional mlo gene had fully developed papillae 24–48 hai as described by Thordal-Christensen et al. [32] and Hückelhoven et al. [14]. The former, complex cell wall aggregation [47] results in decreased reflectance from 420–660 nm and around 710 nm 1–2 dai. This may be explained by cytoplasmic streams in attacked epidermal cells [48]. Similar spectral pattern for Bgh inoculated mlo3 leaves were shown by Kuska et al. [37], already 2 dai. Mla1 based hypersensitive reaction, which cause an early change 12 hai in the NIR range, is correlated with plant physiological and structural changes [37]. This indicates a rapid mode of action within the first 24 hours according to described process of Mla gene based HR [14]. This process seems to decelerate 72 hai, displayed by a turnover of the difference reflectance. In previous studies, this turnover was spectrally determined in a Bgh inoculated Mla1 already 48 hai [37]. This indicates a temporal dynamic of the Mla response against Bgh penetration. Nevertheless, the spectral difference pattern is similar after the HR undergoing cells are necrotized.

Gene expression profiles and their link to hyperspectral reflectance

The LDA analysis on hyperspectral signals or gene expression profiles indicated a high dispersion. This phenomenon derived from the temporal variability during the early individual interactions that depend on the Bgh development and penetration time point [37, 38]. Differences in the separability result from complex physiological and biochemical processes triggering the gene to phenotypic function and are based on the corresponding analysis methods [49]. In contrast, applying the LDA on the combined data sets differentiate all three barley-Bgh interactions independent from the time. This shows significant synergistic effects between the hyperspectral signal and the corresponding gene activities. These results lead to demand for in depth analysis on the temporal variability of the hyperspectral signals and temporarily corresponding gene expression profiles, to elucidate a possible functional link between both highly diverse data sets.

Depending on the barley-Bgh interaction, different genes were expressed. The Bgh infestation begins with the conidia developing an appressorium and starting to penetrate the epidermal cell wall using cell wall degrading enzymes and by an increased turgor pressure [50]. At this time point, PR genes are induced and a down regulation of HvRuBisCO has been described in studies by Eichmann et al. [23]. During the compatible interaction, PR5 showed broad functionality indicated in high frequency of relevant wavebands. PR5 is reported as a marker for systemic acquired resistance, but transcribed proteins are endo-ß-1,3-glucanases as well as endo-ß-1,4-xylanases with anti-fungal activity [19]. A possible xylanase activity is indicated by high relevant wavelengths around 710 nm. Reflectance intensity around 700 nm is characterized by leaf and cell structure [51]. The cell structure is highly influenced by the cell wall in which xylan is a major component for the cell wall of monocotyledonous plants and can be degraded by endo-ß-1,4-xylanases [52]. This coherency indicates that PR5 activity is important during compatible barley-Bgh interaction.

Further important gene activity during compatible barley-Bgh interaction was revealed by upregulated HvGlb1, which is stimulated by gibberellic acid and transcripts of ß-1,3-glucanase mainly in aleuron cells to hydrolyze the cell walls [53]. Previous studies indicate an early up-regulation of ß-1,3-glucanases in Bgh inoculated barley leaves [54]. Similar upregulation profiles are caused by necrotrophic plant pathogens such as Fusarium graminearum infestation on barley roots, which underlines a basal resistance response [55]. Interestingly, a second overexpression of HvGlb1 was during secondary mycelia growth, which initiates a second penetration phase. The highest relevant wavelength bands for HvGlb1 expression profile are in accordance with the mainly influenced wavelengths during early powdery mildew pathogenesis, which reveals HvGlb1 as a possible Bgh penetration indicator [9, 34, 37].

The influence of Bgh on expression of JIP23 was as expected. After Bgh invaginates the epidermal cell and develops a functional haustorium, effector proteins are secreted by Bgh to successfully infest the plant and to overcome plant immunity [56]. Investigation by Scheler et al. [57] provided evidence of early HvJIP23 transcript suppression due to Bgh infestation, which is strongly linked to jasmonate activity and its associated induced systemic resistance [21]. In addition, jasmonates cause loss of chlorophyll, underlining their role in leaf senescence [20]. Relevant wavelength bands around 550 nm, 640–690 nm and 720–830 nm for HvJIP23 expression profiles in Bgh inoculated WT leaves are known to correlate to the chlorophyll content, the photosynthetic activity and the leaf cell structure [51, 58].

HSI reveals the metabolism during incompatible barley - powdery mildew interactions

Upregulated HvDRF and HvRuBisCO in Bgh inoculated mlo3 are evidence for a finished resistance response [59]. This result is in accordance with studies by Swarbrick et al. [22] that analyzed a high transcript amount of RuBisCO related genes in a Bgh inoculated mlo5 genotype from 3 dai. The HvRuBisCO expression profile in this study underlines relevant wavebands in the green and red range, elucidating the coherency of RuBisCO to the photosynthesis apparatus [60, 61]. In addition, wavelength bands in the NIR are relevant due to the influence of RuBisCO on barley leaf cell development [62].

Further spectral coherency of gene expressions in mlo3 were revealed for HvPR3 expression, which is not known to be directly involved in barley mlo base resistance response. Relevant wavelength bands for HvGlb1 expression profiles in Bgh inoculated Mla1 reveals important spectral ranges for early Bgh infestation. If HvGlb1 is a potential Bgh penetration indicator, these results are in accordance with studies by Kuska et al. [37]. They observed several successful penetrations of Bgh on resistant Mla1 leaves that were later stopped by strong HRs. HvPR2 and HvPR5 expression profiles and corresponding relevant wavelengths underline the important role of HvPR gene regulation during HR and cell collapsing against powdery mildew infestation [63].

The link between phenomic data and known parameters from omic and physiological studies is highly demanded to establish high-throughput phenotyping pipelines in crop breeding and crop improvement [4, 64]. Results of the present study successfully demonstrate a link between gene expression and hyperspectral reflectance. According to the hypotheses by Furbank and Tester [2], it is also possible to study the function of genes using the phenomic information derived from the hyperspectral reflectance. The Relief algorithm determined the relevant wavelength bands for a specific gene expression profile and enabled the interpretation of the coherency based on the gene transcript responsibility. Until now, only a few studies have attempted to find these functional relations among spectral phenomics and genomic data. Recently, sugar beet inbred lines differing in two quantitative trait loci against Cercospora leaf spot could be differentiated according to disease severity using HSI [11]. Investigations by Kuska et al., [65] revealed that the combination of invertase analysis and multispectral imaging was a complementary validation system in a high-throughput approach. This provides evidence that spectral technologies bear the potential to be implemented in exhaustive breeding processes, but therefore all specific crop-pathogen interactions must be determined to assess the characteristic spectral properties. In a next step, the impact of proteins, secondary metabolites and hormones on spectral profiles during barley-Bgh interactions need to be investigated. This would reveal a comprehensive phenotyping of a genotype and will support the development of hyperspectral gene maps and improve existing crop physiological modelling e.g. [49]. Then, assessed relevant wavebands of specific gene/protein/metabolite activities need to be validated using knock-out mutants. This will establish novel approaches for plant-pathogen interaction studies and crop resistance breeding.