1.

Introduction

The fungi investigated in this study belong to the same phylum (divisions) named “Ascomycota,” a group whose members are also known as Sac fungi, and have evolved from one common ancestor. Living organisms are biologically classified into different units according to their similarity. A species is often defined as a group of organisms capable of interbreeding and producing fertile offspring. More precise measures are often used, such as similarity of DNA, morphology, or ecological niche. Species that are believed to have the same ancestors are grouped together, and this group is called a genus. Each species may include different isolates (strains) which usually result from one or more mutations. In the hierarchy of the binomial classification system, species is above isolate and below genus.

As far as the genus is concerned, these fungi are grouped into three classes: Fusarium, Colletotrichum, and Verticillium. These fungi are pathogens which attack crops, resulting in fungal diseases, which lead to large economic losses.^1,2 For example, V. ahlia causes a wilt disease in hundreds of species of eudicot plants. Many economically important plants are susceptible, including cotton, tomatoes, potatoes, eggplants, and peppers. Solanaceous crop may be infected at any age by fungi causing the Fusarium wilt and Verticillium wilt, with similar symptoms. The diseases therefore cannot be distinguished based on symptoms alone.

Figures 1(a)–1(c)^3,4 show plants infected by the various fungi genera studied here. Figure 1(a) shows potato tubers infected with F. oxysporum which causes a wilt symptom characterized by browning of the vascular ring and the stem end. In Fig. 1(b) Verticillium wilt symptoms on tomato leaves can be seen. Infected plants usually survive, but both the yields and the fruits may become small, depending on the severity of attack. Figure 1(c) shows a tomato infected with C. coccodes which usually causes serious damage to the fruit.

Fig. 1

Different fungi genera affecting crops. (a) Potato tuber discoloration of the vascular ring, caused by the Fusarium wilt, (b) $v$-shaped lesions on tomato leaves, associated with Verticillium wilt, and (c) C. coccodes causes anthracnose on tomatoes and black dot disease on potatoes.

Early detection of phytopathogens is very important for both successful protection and effective treatment,⁵ thus enhancing the chances of recovery and simultaneously preventing environmental pollution. By standard physiological methods, it is very difficult to differentiate between closely related species and strains^6,7 based on cultivation. In addition, fungi identification by visual and microscopic observations are usually time consuming and not always very specific.⁸

Serological, molecular biology methods are based on differences in the nucleic acid genome of the tested fungi, whereas the immunological method is based on differences in the proteins constructing the fungi and on the availability of specific monoclonal antibodies against the tested fungi. These methods are very sensitive and rapid relative to physiological methods for the identification of pathogens; however, their use is still limited due to their availability for only a small number of fungal pathogens.^9–11 Developing such methods for different fungi strains is complicated, costly, and not always possible. Therefore, the use of these methods is expected to be limited and not available for the screening of large numbers of samples of different fungal strains.

Many studies have indicated the potential of FTIR spectroscopy methods for the detection and identification of microorganisms, especially in food products.^12,13 The vast information already achieved about spectral bands obtained from FTIR spectra of living cells¹⁴ combined with features of infrared spectroscopy such as sensitivity, rapidity, low expense, and simplicity¹⁵ enabled us to differentiate the fungi even on the levels of species^16,17 and strains.^18–23 The use of Fourier transform infrared-attenuated total reflection (FTIR-ATR) spectroscopy followed by principal component analysis (PCA) and linear discriminant analysis (LDA) for biological classification has gained momentum in the last decades (see Martin et al.).^24–26

Linker and Tsror¹⁹ showed good results in differentiating between fungi genera and species using FTIR-ATR and applying statistical analysis techniques: PCA, and cluster and canonical variate analysis (CVA). In a different study, A. Naumann¹⁸ classified 26 fungi strains belonging to 24 different species using FTIR-ATR and applying cluster and artificial neural network (ANN) statistical techniques. Fischer et al.²² tried to develop a method to reproducibly differentiate Aspergillus and Penicillium species on the generic, species, and strain levels. In their work they used nine different species and only one strain from each species. Shapaval et al.²³ showed that it was possible to differentiate 11 species of five different fungal genera using FTIR spectroscopy. Our former study²¹ took the method further into a more challenging stage; by applying the PCA and LDA techniques, the ability of the FTIR-ATR method to classify six different strains of F. Oxysporum was examined. The results showed that it is possible to classify and differentiate between the strains with a 81.4% success rate.

The main goal of this study was to test the feasibility of the FTIR-ATR methodology in differentiating 18 isolates from three different fungi genera: Colletotrichum, Verticillium, and Fusarium. From each of these genera, we selected six isolates originating in a single species, namely six isolates of C. coccodes, six isolates of V. dahliae and six isolates of F. oxysporum. The classification procedure was done in two phases. In the first phase, the fungal samples were classified on the genus level, and in the second phase, the samples were classified on the isolates level. The classification procedures were based on infrared absorption spectra of the samples, and used advanced mathematical and statistical techniques; PCA followed by $k$-means and LDA.

The uniqueness of our study lies in the analysis of a larger number of isolates belonging to the same species. Isolates are very similar; their spectra are blended and overlap, which makes it difficult to differentiate between them, especially relative to the species and genera levels as previously mentioned. Moreover, for pattern recognition methods, increasing the class numbers is a real challenge, especially when the classes are strains of the same species as in this study.

2.

Materials and Methods

3.

Results

Figure 2 shows the infrared absorption spectra of the three fungi genera investigated in this study, i.e., Colletotrichum, Verticillium, and Fusarium, each with a different color. The spectra for each fungus were measured from six different isolates. The main spectral features in the high wavenumber region ($2700–2950{\mathrm{cm}}^{-1}$) are the bands detected at 2853 and $2922{\mathrm{cm}}^{-1}$. These bands derive mainly from phospholipid absorbance.³⁸ Water absorbance bands in this region were excluded from the spectra as part of the analysis procedure. The main features in the low wavenumber region ($800–1775{\mathrm{cm}}^{-1}$), were the amide I and amide II with centroids at $1650{\mathrm{cm}}^{-1}$ and $1553{\mathrm{cm}}^{-1}$, respectively. There is a clear peak which arises from lipids absorbance³⁹ at $1743{\mathrm{cm}}^{-1}$. A large absorbance band at $1076{\mathrm{cm}}^{-1}$ was mainly attributed to carbohydrate and nucleic acid vibrations. The centroid of the amid III band was detected at $1252{\mathrm{cm}}^{-1}$. The chitin band, which is specific to fungi, was detected at $1151{\mathrm{cm}}^{-1}$ and $1078{\mathrm{cm}}^{-1}$ due to its C−O and C−C stretching vibrations, and the glycogen C−O stretching vibration was detected at $1024{\mathrm{cm}}^{-1}$.

Fig. 2

Infrared absorption spectra of C. coccodes, V. dahliae, and F. oxysporum in the ranges 890 to $1770{\mathrm{cm}}^{-1}$ and 2800 to $2990{\mathrm{cm}}^{-1}$.

As previously mentioned, there are clear differences between the fungi mainly in the lower wavenumber region, namely, in the amide I and II, chitin, and glycogen peaks. These differences enabled us to differentiate the three genera using unsupervised methods, such as $k$-means, achieving an accuracy of 94.5%.

As shown in Fig. 2, the Colletotrichum fungus has low intensities in the amide I and II regions which may indicate lower protein content relative to Fusarim and Verticillium fungi.

Utilizing the PCA and LDA (20 to 80) methods, we classified the sample into three groups with great success. When using one PC we obtained an accuracy of 61.2%, using two PCs a 96.9% accuracy was achieved, which increased to 99.7% with three PCs.

Figure 3 shows the results for the three fungal groups obtained by LDA based on three PCs of PCA at the lower wavenumber region.

Fig. 3

Complete date sets of Colletotrichum, Verticillium, and Fusarium presented in a 3-D PCs domain. This figure summarizes the first phase (genus) classification results.

In this figure each spectrum was calculated as a superposition of three loadings (PCs) derived using the PCA algorithm. Each spectrum was identified by three numbers, which are the coefficients of the loadings. As shown in Fig. 3 the data is divided into three groups, red squares (Colletotrichum), blue crosses (Fusarium), and green open triangles (Verticillium). There is an excellent separation between the three groups. To find out which feature contributed to the classification procedure, we plotted two-dimensional figures for PC1 versus PC2, depicted in Fig. 4(a), which provided the best classification between the three studied genera. The projection of the data on the PC1 axis provided a good classification of Fusarium. The projection of the Colletotrichum and Verticillium data on the PC2 axis provided a complete classification between them. The first two PCs are plotted in Fig. 4(b) where the main peaks are labeled. These are: for PC1, the $\mathrm{C}=\mathrm{O}$ stretching mode of lipids ($1743{\mathrm{cm}}^{-1}$), the amide I and II bands ($1641{\mathrm{cm}}^{-1}$ and $1544{\mathrm{cm}}^{-1}$, respectively), the asymmetric and symmetric ${\mathrm{CH}}_3$ bending modes of the protein methyl groups⁴⁰ ($1453{\mathrm{cm}}^{-1}$ and $1397{\mathrm{cm}}^{-1}$), collagen, the asymmetric stretching of phosphate groups of phosphodiester linkages in DNA and RNA⁴¹ ($1224{\mathrm{cm}}^{-1}$), carbohydrates⁴² ($1148{\mathrm{cm}}^{-1}$), the symmetric phosphate ${\mathrm{PO}}_2^{-}$ stretching⁴³ ($1076{\mathrm{cm}}^{-1}$), and ${\mathrm{OCH}}_3$ [polysaccharides-cellulose⁴⁴ ($989{\mathrm{cm}}^{-1}$)]. All the peaks labeled with dashed lines are also common for PC2. The other main peaks of PC2 are: C−N stretching and C−N−H bending of N−H⁴⁵ ($1478{\mathrm{cm}}^{-1}$), collagen⁴⁶ ($1345{\mathrm{cm}}^{-1}$), amide III, collagen⁴⁷ ($1280{\mathrm{cm}}^{-1}$), phosphate, oligosaccharides,⁴⁸ and carbohydrates⁴² ($1146{\mathrm{cm}}^{-1}$), stretching C−O ribose⁴⁹ ($1068{\mathrm{cm}}^{-1}$), C−C [Ref. 49($994{\mathrm{cm}}^{-1}$)], and the phosphodiester stretching bands region [for absorbance due to collagen and glycogen⁵⁰ ($926{\mathrm{cm}}^{-1}$)].

Fig. 4

(a) Two-dimensional plot of Colletotrichum, Verticillium, and Fusarium included in this study. (b) This figure presents the first 2 PCs, derived from PCA, in the lower range. The main peaks of each PC are labeled.

In the second phase we tried to classify the six different isolates within each group (genus). In contrast to the genus level where there were clear differences between the spectra, here the spectra of the different strains were blended and overlap [Figs. 5(a)–5(c)], and in each spectrum all major absorption bands described earlier were detected. However, some differences were found, such as different absorbance intensities in different absorbance bands in the two regions.

Fig. 5

Infrared absorption spectra of six isolates of C. coccodes (a), six isolates of V. dahliae (b), and six isolates of F. oxysporum (c) in the ranges 800 to $1770{\mathrm{cm}}^{-1}$ and 2800 to $2990{\mathrm{cm}}^{-1}$.

The analysis was performed for different regions of the spectrum, and the best results were achieved when the lower wavenumber region ($800–1775{\mathrm{cm}}^{-1}$) was used. The $k$-means achieved poor results, thus we adopted the PCA and LDA methods. In the second phase where the number of measurements for each strain was low, the “leave-one-out” method was preferred, whereas in the first phase where the statistics were good the 20 to 80 algorithm was used.

Figure 6 shows the identification errors (percentage) derived using LDA together with the “leave-one-out” and 20 to 80 algorithms, as a function of the PC number used in the analysis of Colletotrichum. Due to the similarity between the strains of each of the three groups, more PCs had to be used than in the first (genus) phase.²¹ Nine PCs were used in the strain phase, which achieved a good classification and simultaneously kept the highest loading (PC) meaningful and noiseless.

Fig. 6

Identification error vs. PC No. in the lower wavenumber region.

The LDA results using the “leave-one-out” algorithm and nine PCs showed that it was possible to differentiate between the Colletotrichum isolates with a 86.8% success rate, whereas the classification success rates of the Fusarium and Verticillium strains were 85.5% and 94.6%, respectively. The LDA results using the 20 to 80 algorithm and 9 PCs showed that it was possible to differentiate between the Colletotrichum isolates with a 84.6% success rate, whereas the classification success rates of the Fusarium and Verticillium strains were 80.7% and 94.3%, respectively.

Table 1 shows the results of the classification of the studied genera on the isolate level using the “leave-one-out” method.

Table 1

Successful identification of (a) Colletotrichum (Coll), (b) Fusarium (Fus) and (c) Verticillium (Vert) isolates obtained using LDA calculations and the “leave-one-out” algorithm in the low wavenumber region (900–1775  cm−1).

4.

Discussion

In principle, the ATR sampling technique is similar to measurements done using remote fiberoptic probes. Thus, evaluation of the ATR potential for fungal classification on large numbers of strains on the levels of genus, species, and isolates is an important step toward in vivo measurements using infrared fibers.

Many studies have shown that it is possible to use FTIR-ATR for detection and identification of fungi.^18,19,21 In their study, Linker et al.¹⁹ managed to classify different fungi genera species and to differentiate between two isolates of Colletotrichum based on the measured FTIR-ATR spectra and advanced statistical methods: CVA and PCA. Based on the ATR sampling technique, Naumann¹⁸ succeeded to differentiate between 26 fungal isolates that belong to 24 different species using cluster analysis and ANN analysis.¹⁸

In their work, Fischer et al.₂₂ used nine different species and only one strain from each species. Shapaval et al.²³ used 11 different species of five different fungal genera. In our previous work, we used the FTIR-ATR to differentiate between six different isolates of F. oxysporum.²¹

By enlarging the number of strains, we come closer to reality, where tens of isolates from each species exist. All previous studies focused on just a few species belonging to the same genus, and a few strains belonging to the same species. Thus, success in the classification of large numbers of isolates and species provides a solid base for the future large scale in vivo examination of phytofungal pathogens using infrared spectroscopy.

In this study we investigated six different isolates from each of three fungi genera Colletotrichum, Verticillium and Fusarium. Our main objective was to test and evaluate the potential of FTIR-ATR spectroscopy to differentiate between these isolates. This was done in two phases of examinations of fungi samples, the first on the genus level and the second on the isolate level. The fungal isolates to be tested were collected, isolated, and purified to obtain pure samples of the desired strain.

Our analysis was based on the fungal IR evanescent wave absorption spectra. We used a statistical approach to analyze the large database of the obtained IR spectra.

Using the currently practiced classic physiological methods, it is very difficult to accurately differentiate between the strains because of their high morphological similarity. Molecular and serological methods are also not available for all these strains.

Due to the variance in the penetration depth occurring in the ATR mode measurements, absorption band intensities and positions are different than those obtained using the transmission mode.^51–53 Thus, all our spectra have undergone the same penetration depth correction using OPUS software. On the genus level, there were clear differences that have enabled the differentiation between the samples by unsupervised clustering analysis using $K$-means, with a 94.5% success rate. In-house developed software codes based on MATLAB software⁵⁴ were used. After PCA calculation, a supervised classifier LDA was applied, which enabled the classification of the data with an accuracy of 99.7%. In this phase of the work (genus level), 3 PCs were sufficient to obtain such high accuracy, whereas in the second phase (isolate level) nine PCs were required. PCA enabled capturing the variability of the fungi using a small number of PCs. In the first phase the examined fungi genera were successfully differentiated with a high accuracy using two PCs, and using three PCs obtained almost a 100% success rate. The second phase, in which the samples were isolates of the same species, the differences between their spectra were minor, and therefore a good differentiation was achieved using a larger number of PCs. It is not guaranteed that the first PCs are the most important for separability; however, they carry most of the variance of the data.

The dependence on the PC number is task-dependent, i.e., number of classes to classify, level of classification (genus, species, or isolates), and also the type of samples.

From Fig. 4(a) we can see that the new directions defined by PC1 and PC2 give a good classification. Using PCA calculation, we obtained a new basis which defines new directions.⁵⁵ Figure 4(b) shows PC1 and PC2 vectors in the standard basis. The peaks which are found in PC2 and not in PC1 (labeled by dashed lines) are suspected to be more dominant with respect to the classification. The comparison procedure of the peaks is relatively easy when a small number (2 to 3) of the PCs have to be compared.

Smaller numbers of PCs make it easier to relate the changes of the PCs shape to biology, whereas this is much harder when a larger number of PCs is used. In this study, the goal was classification, and thus we should choose the number of PCs which provided the best classification providing that the highest PC was still meaningful for the biological system investigated. This choice of PC number is still not objective in our field of spectroscopy and should be investigated more carefully.

As shown in Fig. 4(a), a good classification of Fusarium was achieved. This is probably due to some peaks in PC1 which are absent in PC2 [Fig. 4(b)]. These peaks which contribute mainly to the classification of Fusarium are $1453{\mathrm{cm}}^{-1}$ (methyl groups of proteins),⁴⁰ $1224{\mathrm{cm}}^{-1}$ (collagen, asymmetric, stretching of phosphate groups of phosphodiester linkages in DNA and RNA),⁴¹ $1148{\mathrm{cm}}^{-1}$ (chitin, carbohydrates)⁴² and $1076{\mathrm{cm}}^{-1}$ (chitin, symmetric phosphate ${\mathrm{PO}}_2^{-}$ stretching).⁴³

PC2 gives a good classification between Colletotrichum and Verticillium. It can be seen in Fig. 4(b) that some peaks in PC2 do not exist in PC1, and this is the main contribution to the classification between Colletotrichum and Verticillium. These peaks are $1478{\mathrm{cm}}^{-1}$ (C−N stretching and C−N−H bending N−H),⁴⁵ $1345{\mathrm{cm}}^{-1}$ (collagen),⁴⁶ $1280{\mathrm{cm}}^{-1}$ (amide III, collagen),⁴⁷ $1146{\mathrm{cm}}^{-1}$ (phosphate, oligosaccharides,⁴⁸ and carbohydrates)⁴² and $1068{\mathrm{cm}}^{-1}$ (C−O ribose),⁴⁹ $994{\mathrm{cm}}^{-1}$ (C−O ribose, C−C)^,49 and $926{\mathrm{cm}}^{-1}$ [phosphodiester stretching bands region (for absorbance due to collagen and glycogen)].⁵⁰

The LDA calculation was designed applying the two methods (“leave-one-out” and 20 to 80), using different sets for prediction each time, so that the test sets would be statistically independent. The “leave-one-out” method is a common method of cross-validation, extensively explored in machine learning, used to estimate the error in a small sized populations. Using it ensures the validation of results.^56,57

The 20 to 80 method is used when large numbers of data exist, thus it was used in the first phase of the study. In the second phase of identification the “leave-one-out” was preferred because it achieved better results, but the 20 to 80 was used as well to validate the “leave-one-out” algorithm (Fig. 6).

Our results with a successful classification of the different strains of each genus using the “leave-one-out” method, are presented in Table 1(a)–(c).

In the second phase, nine PCs were used, thus the discrimination power is the overall contribution of the whole set of PCs. It is difficult to outline which PC contributes more to the discrimination between the different isolates (strains).

It is very difficult to discuss the relevance of spectral differences of the various species or isolates, due to the high similarity in the molecular and structural composition of these related isolates, which belong to the same species. According to the FTIR-ATR spectra, the differences are spread over the entire spectrum and not specific to a certain band. One limitation of FTIR is that the spectra obtained from biological samples are complex, and difficult to interpret because it is hard to know which bands arise from which biomolecule. For example, changes in the membrane are reflected in the absorption bands of lipids; however, spectral differences in the lipid absorption bands may not arise solely from membranes, but may be attributed to lipids that are not located in the cell membrane.

The applied method is objective and computerized. Enlarging the database and the strain number will improve the statistics and bring the method closer to reality where large numbers of strains exist. The method shows a great potential for the identification and study of phytofungal pathogens, as far as the level of strain identification.

This method may be useful for studying biological aspects of the different genera as shown in Fig. 2. It shows significantly lower intensities in the amide I and II regions for Colletotrichum compared with Fusarim and Verticillium fungi. These differences may reflect different protein contents between these genera. Certainly, small differences in their structural components (such as lipids, proteins, and sugars) do exist, but unfortunately the origins of these differences are yet unknown and it is very difficult to define the exact biological origin of the observed spectral differences. However, these difficulties do not affect the main objective of this study, which was to examine the potential of spectroscopic techniques for reliable detection and discrimination between different fungi strains.

5.

Conclusions

FTIR-ATR spectroscopy in tandem with PCA methods, followed by LDA calculations, enabled the differentiation between the fungal samples studied not only on the genus level but on the level of strains as well (where spectral differences are minute) with a good confidence level.

Enlarging the database to include more species and strains, could improve the statistics and bring the method one step forward toward real conditions, where tens of isolates from the same species exist.