Investigation of diabetes-induced effect on apex of rat heart myocardium by using cluster analysis and neural network approach : An FTIR study

Diabetes mellitus (DM) is a progressive chronic disorder, which affects people belonging to all age groups of the population. This disease is accompanied by a greatly increased risk of cardiovascular death. In the present study, the effects of streptozotocin (STZ)-induced type 1 diabetes on apex myocardium of the rat heart have been investigated using Fourier Transform Infrared (FTIR) spectra. The cluster analysis has been applied to FTIR spectra to differentiate the diabetic samples from the normal controls. In addition, the protein secondary structures of diabetic and normal tissues were predicted by neural networks based on the amide I band of the FTIR spectra. The findings mainly suggest that 5 weeks of diabetes alters the lipid and protein profile of normal rat heart apex myocardium, which might have an important role in understanding the molecular mechanism of diabetes-related heart diseases.


Introduction
Diabetes Mellitus (DM) became a serious health problem throughout the world.World Health Organization estimated that the worldwide prevalence of diabetes is expected to grow from 171 million in 2000 to 366 million by 2030 [1].There are two types of this disease namely, type 1 and type 2. Type 1 occurs due to the loss of insulin production in the beta cells of the pancreas, while type 2 occurs due to lack of serum insulin or poor uptake of glucose into the cells [2].Both types of diabetes are known to affect normal heart function, which eventually leads to congestive heart failure even in the absence of coronary artery disease [3].In diabetes, hyperglycemia causes rapid changes in membrane function, followed by contractile dysfunction within weeks [4][5][6].Cardiomyopathy in diabetes is associated with decreased diastolic compliance, interstitial fibrosis and myocyte hypertrophy, to mention a few.The molecular mechanisms leading to this disease are still uncertain.
Previously, the effect of diabetes have been documented in many studies, including our's, on different regions of the heart, such as left ventricle myocardium [7][8][9], right venticle myocardium [7,10,11], papillary muscles [13,14], and ventricular myocytes [13].However, the number of studies investigating the effect of diabetes on the apex myocardium is very limited.In one of them, the effects of diabetic cardiomyopathy have been investigated on the rat apex with respect to electrophysiological characteristics [15].Apex is a clinically important region of the heart in terms of both diagnosis and treatment of heart diseases [16][17][18][19][20].For instance at cardiac apex, a mitral systolic murmer is audible as a holosystolic or late systolic murmer [21].Molecular changes in the cardiac apex are known to be very important for the regulation of the normal electrical activity of the heart [22].Therefore, in the present study, we have investigated the effects of type 1 diabetes on the protein secondary structure of apex myocardium of the rat heart using neural networks based on Fourier Transform Infrared (FTIR) spectra to understand the underlying biochemical changes associated with diabetes.In addition, we have differentiated successfully the diabetic tissues from the control ones with cluster analysis for investigating the changes in lipids (at 2800-3050 cm −1 spectral region), and in proteins (1480-1800 cm −1 spectral region) using FTIR data.

Materials and methods
The control and the diabetic groups of the Wistar rats used in this study were prepared as described previously [7].Two adjacent cross-sections of tissues having 9 µm thickness were taken from the apex of all groups for acquisition of spectral data.In order to record FTIR spectra, these sections were thawmounted on IR-transparent CaF 2 windows.The rat heart was attached to the cryotome by using a small amount of optimum cutting tool.The first tissue sections were used for FTIR microspectroscopy measurements and the second serial sections were used for Hematoxylin & Eosin (H&E) staining.This staining was performed to see the histologically defined tissue regions on the sections.To map the tissue sections, an FTIR microspectrometer (Bruker, Germany) was used.The details of this IR mapping are described previously [7].To have IR data completely covering the tissue area of interest, spectra were collected in both x and y directions in steps of 80 µm.IR absorption spectra were collected from 850 to 4000 cm −1 .The resolution was 6 cm −1 .The number of scans co-added per pixel spectrum was 64.The spectrometer was continuously purged with dry air to get rid of spectral contributions from water vapor and CO 2 .

Spectral analysis
For microspectroscopic data analysis, CytoSpec (http://www.cytospec.com)and OPUS data collection software packages (Bruker, Germany) were used.Quality test was applied to the raw spectral data prior to the data evaluation procedure, as described previously [7].The spectra which failed the quality test were removed and were not used for further analysis.The remaining spectra were used for first derivative calculations in 950-1480 cm −1 and 2800-3050 cm −1 spectral regions, using a five smoothing point Savitzky-Golay algorithm.The next step was the vector-normalization.For this purpose, first derivative spectra in the frequency range of 950-1480 cm −1 were used.Spectral classes corresponding to specific tissue structures were identified using cluster analysis-based maps.For this purpose, cluster analysis was performed in 950-1480 cm −1 and 2800-3050 cm −1 regions by using first derivative spectra to obtain the average spectra arising only from the myocardium of the rat cardiac apex.The original absorption spectra and their averages belonging to different clusters were saved.For further analysis, the data were loaded into OPUS.

Cluster analysis
To facilitate comparison of normal and diabetic cardiac apex myocardia, cluster analysis were performed on second derivative spectra using a nine smoothing point Savitzky-Golay algorithm in 2800-3050 cm −1 and 1480-1800 cm −1 spectral regions for the analysis of signals arising from lipids and proteins, respectively.Spectral distances were calculated between pairs of spectra as Pearson's correlation coefficients [23].For separation of control and diabetic tissue, cluster analysis was based on the Euclidean distances.Ward's algorithm was used for hierarchical clustering in all cases.The hierarchical clustering is a multivariate statistical data analysis method, which builds a hierarchy of clusters from individual elements.With cluster analysis, it is possible to examine the interpoint distances between all the samples and represents the information in the form of a two dimensional plot.This plot is known as dendogram, which is a tree-like diagram demonstrating the arrangement of the clusters.Each subset of spectra having closer similarities to each other than another set of spectra are classified in a single cluster [24].Even small differences between groups can be visualized easily by cluster analysis.

Neural network analysis
We have used neural networks to predict the secondary structure content of proteins.The neural networks were initially trained using a data set containing FTIR spectra of 18 water soluble proteins recorded in water using the method described in reference [25].The secondary structures of these proteins were known from X-ray crystallographic analysis.The size of the data set was increased by interpolating the available FTIR spectra to improve the training of the neural networks.Before training, amide I band located at 1600-1700 cm −1 of FTIR spectra were first normalized and their discrete cosine transforms (DCT) were obtained.To improve the generalization property of the neural network, the number of inputs were restricted to the significant DCT coefficients.Bayesian regulation was used to train the neural networks whose structures were optimized in terms of the number of inputs and number of hidden units.The trained neural networks have standard error of prediction values of 4.19% for αhelix and 3.49% for β-sheet.Although these errors may seem to be large, it should be pointed out that neural networks yield more reliable results for incremental changes in α-helix and β-sheet contents as in this study.The details of the training and testing algorithm can be found in [25].

Statistical analysis
The results were expressed as "mean ± standard error of mean".Mann-Whitney U test was used to test the significance of the differences between control and diabetic groups.p values less than 0.05 were accepted as significantly different from the control group.The degree of significance was denoted as * p < 0.05.

Results and discussion
Due to the increasing importance of diabetes-related cardiovascular diseases, in the current study, we have investigated the effect of diabetes on rat heart apex myocardium by using cluster analysis for comparison of spectral changes in lipids and proteins between control and diabetic groups, and by neural networks for protein secondary structure estimation.
Figure 1(a) demonstrates the H&E staining image of a section with 9 µm thickness taken from the rat heart apex at ×25 magnification.The mapped region is shown in square on this section.After completing FTIR microscopy measurements, the data were loaded to cytospec program for data analysis procedure.In the cluster imaging of IR, clusters should contain spectra from histological regions exhibiting similar spectral features.Spectra in different clusters ideally manifest different spectral signatures.Assigning a distinct color to all spectra in one cluster is the main idea in image assembly on the basis of cluster analysis [26].The image of cluster analysis, which was performed to discriminate myocardium of the apex, is shown in Fig. 1(b), (c) illustrates the average spectra arising from different clusters.In Fig. 1(b) and (c), it is possible to see the 4 different clusters given in gray scale ranging from white color (numbered as 2 in Fig. 1(b)) to very dark gray color (numbered as 3 in Fig. 1(b)) belonging to different components, and the original average absorption spectra of these clusters, respectively.The cluster shown by dark gray color, numbered as 4 in Fig. 1(b), arises from the tissue freezing medium (optimum cutting tool) and the spectrum belonging to this cluster is numbered as 4 in Fig. 1(c).The cluster shown by very dark gray color, numbered as 3 in Fig. 1(b), arises from the epicardium of the apex and the tissue freezing medium.The corresponding original average absorption spectrum is also illustrated with the same number in Fig. 1(c).The cluster represented by white color, numbered as 2 in Fig. 1(b), belongs to the epicardium of the apex.The cluster shown with gray color, numbered as 1, arises from the myocardium of the apex and the same number is used to show the original average absorption spectrum in Fig. 1(c).The spectra which failed the quality test were excluded from all consecutive evaluations and are shown with black color in Fig. 1(b).The aim of this cluster analysis is to obtain the average original absorption spectra arising only from the myocardium of the apex since we are interested in the changes occurring in this specific region of the cardiac apex between control and diabetic groups.The data belonging only to this cluster showing the apex myocardium is used for the comparisons.The cluster analysis was performed for the apex part of all the rat hearts and the original average spectrum arising from myocardium was saved.For further analysis, the data were loaded into OPUS program.Then, cluster analysis was done for comparison of normal and diabetic groups for apex myocardium of the rat heart.Representative average absorption spectra belonging to control apex myocardium is given in Fig. (2).In Fig. 2(a), absorption spectrum of the control apex myocardium is given in 3050-2800 cm −1 region.Figure 2 The major absorptions in 3050-1480 cm −1 are numbered in this figure and the frequency values with their assignments are given in Table 1.
The spectral information contained in the second derivative spectra was used as input data for hierarchical cluster analysis in order to obtain objective classification on the basis of spectral patterns.Cluster analysis results are displayed as dendograms using spectral information between 2800 and 3050 cm −1 ; and using spectral information between 1480 and 1800 cm −1 of normal and diabetic groups for apex myocardium in Figs 3 and 4, respectively.The 2800-3050 cm −1 region mainly contains C-H stretching vibrations arising from the fatty acyl chains of membrane lipids, and the spectral region rich in proteins is between 1800 and 1400 cm −1 [27,28].As seen from the figures, cluster analysis resulted into two distinct clusters corresponding to control and diabetic groups with success of 6 out of 6 and 8 out of 8, respectively, in both spectral regions subjected to this analysis.These results imply that diabetes causes some important alterations not only in lipids but also in proteins of the apex myocardium.These findings of this study demonstrated the feasibility of FTIR microspectroscopic discrimination of diabetic and nor- Amid II: (protein N-H bending, C-N stretching), α-helix Fig. 3. Cluster analysis results displayed as dendograms using spectral information between 2800 and 3050 cm −1 of normal and diabetic groups for apex myocardium.
Fig. 4. Cluster analysis results displayed as dendograms using spectral information between 1480 and 1800 cm −1 of normal and diabetic groups for apex myocardium.
mal cardiac tissues.The diabetes-induced changes in the lipids might be due to the altered myocardial energy substrate utilization, which consequently may have a detrimental effect on heart function due to accumulation of lipid intermediates [29].
The protein region (amide I), corresponding to absorption values between 1600 and 1700 cm −1 , was further analyzed to determine the diabetes-induced changes in the protein secondary structure using neural network predictions based on FTIR data.A number of techniques, ranging from simple tissue staining to highly specialized methods including X-ray crystallography, NMR and protein sequencing can be used for the identification of molecules associated with DM.While performing conventional analytical procedures, some specific property in a sample might be damaged or denatured due to extraction, fixation or staining.Specifically for protein structure determinations, conventional procedures used for isolation of proteins are too destructive to derive accurate information about the real structure of proteins.With FTIR spectroscopy, it is possible to monitor molecular and structural composition directly in the untreated and unfixed whole tissue [30].IR spectroscopy opens a new field of medical research, as it causes no damage to the important constituents of the cells or tissues [31].In addition, recent developments in data processing techniques and instrumentation made IR spectroscopy a useful tool in the medical arena.Furthermore, having objective results is another important advantage of using IR spectroscopy technique, since spectral data are collected and treated by computer controlled algorithms.
Recently, this technique has successfully been applied for the determination of protein secondary structure in solution [32], in tissues [33] and in membranes [34].The amide I region (1700-1600 cm −1 ) is commonly used for the analysis of secondary structure of proteins in FTIR spectra.In this particular spectral region, different protein conformations result in different discrete bands, which are usually broad and consequently overlapping.For the identification of the bands seen in FTIR spectra, mathematical resolution enhancement techniques are commonly used [35][36][37].Curve fitting technique is one of them which was originally used by Byler and Susi [35], to analyze protein amide I bands.One of the main disadvantages of the curve fitting procedure is that it requires a series of subjective decisions, such as assignment of peaks, which can significantly alter the results [38][39][40].Another disadvantage is that curve fitting has a tendency to overestimate the beta sheet content of the primarily helical proteins.These disadvantages of curve fitting approach stimulated the development of new methods such as neural networks.This new computational technique has proven to be an alternative powerful tool for the analysis of protein structure from FTIR spectra [41].
The neural network prediction results are presented in Table 2.It is clearly seen from the table that diabetes causes significant changes in the protein secondary structure of cardiac apex myocardium by decreasing the content of α-helix from 69.88 ± 2.34 to 48.97 ± 5.42 ( * p < 0.05), and by increasing the content of β-sheet structures from 11.70 ± 2.83 to 43.12 ± 9.03 ( * p < 0.05).So, we can deduce that DM is altering the secondary structure of proteins which might be indicative of either a structural rearrangement of already existing proteins or the expression of new types of proteins having different structural compositions.The conformational changes we have observed in the secondary structure of proteins might also be due to cleavage of proteins in diabetes.As suggested by Kugiyama et al. [42], proteins exposed to glucose are cleaved and undergo conformational changes.These changes were shown to be dependent on hydroxyl radicals, which might be produced by glucose auto-oxidation [43].In previous studies, myocardial cell structure was reported to be destructed due to hyperglycemia [44], and cardiac structure was reported to be changed in diabetes [29].Although no detailed secondary structure analysis has been performed previously, diabetes-induced conformational changes were indicated from the changes in the amide I band shape in rat heart homogenates [45].Another possible explanation for the changes in the protein structure might be the impaired heat shock protein (Hsp) response, which was previously reported to occur in diabetes [46].In the current study, we obtained significant insights into the changes in protein secondary structure of cardiac apex myocardium, which might be important in understanding the molecular mechanism lying behind diabetes-induced heart diseases.With the technique we have used in the current study, we were able to monitor changes in the secondary structure of proteins in tissues without isolating them.Membrane proteins are difficult to be isolated and only a very limited number of proteins were so far isolated.It is crucial to reveal diabetes induced alterations for the development of new drugs to at least treat the diabetes induced damages at molecular level.

Conclusions
In this study, application of cluster analysis to FTIR spectra permitted a rapid and reliable discrimination between control and diabetic group in both lipid and protein regions.Furthermore, neural network approach based on FTIR data has been used for the first time in this study to reveal diabetes induced changes in the secondary structure of proteins in apex myocardium of the rat heart.Present study points out the importance of FTIR microspectroscopy as an excellent technique for the estimation of protein secondary structure.

Fig. 1 .
Fig. 1.Light microscope image of an H&E-stained section taken from the apex of rat heart at ×25 magnification including the mapped region shown in a square (a), image of cluster analysis (b) and average spectra belonging to different clusters (c). 4 different clusters are given with different colors and numbers ranging from 1 to 4 in (b), and the corresponding original average absorption spectra are illustrated with the corresponding numbers in (c).

Table 1
[7]d assignments of major absorptions in IR spectra of control cardiac apex myocardium in 3050-1480 cm −1 spectral region[7]

Table 2
The results of neural network predictions based on FTIR data in 1600-1700 cm −1 spectral region for the changes in protein secondary structure between control and diabetic groups Data shown as mean ± standard error of mean.P < 0.05 were accepted as significantly different from the control group.The degree of significance was denoted as: * p < 0.05.