Toward a Spectroscopic Hemogram: Raman Spectroscopic

May 25, 2012 - Institute of Photonic Technology, Jena, Germany. ∥. Institute for Physical Chemistry and Abbe Center of Photonics, Friedrich Schiller...
2 downloads 0 Views 3MB Size
Article pubs.acs.org/ac

Toward a Spectroscopic Hemogram: Raman Spectroscopic Differentiation of the Two Most Abundant Leukocytes from Peripheral Blood Anuradha Ramoji,†,‡ Ute Neugebauer,*,†,‡ Thomas Bocklitz,∥ Martin Foerster,§ Michael Kiehntopf,†,⊥ Michael Bauer,† and Jürgen Popp†,∥,‡ †

Center for Sepsis Control and Care (CSCC), Jena University Hospital, Jena, Germany Institute of Photonic Technology, Jena, Germany ∥ Institute for Physical Chemistry and Abbe Center of Photonics, Friedrich Schiller University, Jena, Germany § Internal Medicine I, Pneumology, University Hospital Jena, Jena, Germany ⊥ Institute for Clinical Chemistry and Laboratory Diagnostics, Jena University Hospital, Jena, Germany ‡

ABSTRACT: The first response to infection in the blood is mediated by leukocytes. As a result crucial information can be gained from a hemogram. Conventional methods such as blood smears and automated sorting procedures are not capable of recording detailed biochemical information of the different leukocytes. In this study, Raman spectroscopy has been applied to investigate the differences between the leukocyte subtypes which have been obtained from healthy donors. Raman imaging was able to visualize the same morphological features as standard staining methods without the need of any label. Unsupervised statistical methods such as principal component analysis and hierarchical cluster analysis were able to separate Raman spectra of the two most abundant leukocytes, the neutrophils and lymphocytes (with a special focus on CD4+ T-lymphocytes). For the same cells a classification model was built to allow an automated Raman-based differentiation of the cell type in the future. The classification model could achieve an accuracy of 94% in the validation step and could predict the identity of unknown cells from a completely different donor with an accuracy of 81% when using single spectra and with an accuracy of 97% when using the majority vote from all individual spectra of the cell. This marks a promising step toward automated Raman spectroscopic blood analysis which holds the potential not only to assign the numbers of the cells but also to yield important biochemical information. lood is an important body fluid which contains three major corpuscular elements: red blood cells (erythrocytes), white blood cells (leukocytes), and platelets (thrombocytes). These particles, the solid portion of the blood, make up totally 45% of the blood volume. Scheme 1 shows how the corpuscular blood elements can be divided into the different types. The leukocytes can be further divided into major subtypes based on the shape of their nucleus: the polymorphonuclear granulocytes (neutrophils, eosinophils, and basophils) which possess a multilobed nucleus, and the mononuclear monocytes and lymphocytes. Neutrophils and lymphocytes are the most common leukocytes and form together almost 90% of the leukocyte population. Both of them are very important in combating infections.1 Neutrophils have a diameter of 10−12 μm and possess a nucleus with two to four lobes. Hence, they are also known as polymorphonuclear neutrophils (or PMNs). They belong to the granulocyte family, i.e. they possess cytoplasmic granules containing oxidative metabolites and digestive enzymes. When they are in contact with foreign agents or microorganisms neutrophils undergo rapid metabolic and morphological changes. The main functions of neutrophils are phagocytosis

B

© 2012 American Chemical Society

and digestion of bacteria and other pathogens. In this context the formation of neutrophil extracellular traps (NETs) has been discussed recently.2−5 Lymphocytes can be further subdivided into natural killer cells, B lymphocytes and T-lymphocytes, which can be even further subclassified. With only 7−9 μm in diameter Tlymphocytes are smallest in size compared to other leukocytes. They are mononuclear cells (MNCs) with a single large nucleus and a narrow cytoplasmic fringe. Number, size, morphology, and chemical state of the blood leukocytes in the body reflect very well the response of a patient to a disease process or adverse environmental conditions. Therefore, recording of a hemogram is part of everyday routine in every hospital to assist medical diagnosis. Classically, different staining solutions are used to visualize the morphological differences between those cell types. In modern analyzers and sorting machines specific antibodies and dyes are Received: March 19, 2012 Accepted: May 25, 2012 Published: May 25, 2012 5335

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342

Analytical Chemistry

Article

Scheme 1. Schematic Tree Depicting the Types of Corpuscular Elements in the Blood and Their Major Subpopulationsa

a

The percentages state the relative abundances for healthy persons.

used to assist the identification. The routine hemogram records numbers of the different corpuscular components. Depending on the used automates further information on size, morphology, and granularity can be gained. Information on the activation state and biochemical changes in the cells is not detected. However, as those activation-dependent changes in the leukocytes can be different depending on the underlying condition of the patient (e.g., infections, chronic inflammation, and cancer)6 important information that could be utilized for effective diagnostic and therapeutic decision making is lost in the routine hemogram. Therefore, a reliable and fast diagnostic method which not only provides the identification and differentiation of the cells but can also give deeper insights on the biochemical changes without the preselection by special labels is of high interest. Vibrational spectroscopy is a label-free, nondestructive, and chemically very specific method that could address those questions and has been employed in earlier studies to successfully investigate leukocytes by means of Raman7−25 and infrared absorption spectroscopy.26−30 Raman-based studies have investigated specific cellular activities or characteristics and visualized different cellular components within leukocytes also during activation.8,9,18,19,21,25,31−34 Activation studies focused on lymphocytes, neutrophils, and eosinophils and could prove that ex-vivo activation and renal allograft rejection causes specific changes in the leukocytes that can be identified by Raman spectroscopy. Furthermore, the specificity of the Raman spectral fingerprint region could be used to differentiate between leukocytes and other cell types, such as tumor cells.13−15 A systematic study to differentiate leukocyte subpopulation from human blood by means of Raman spectroscopy without the need of any labels has not been reported so far, but is the necessary starting point for a spectroscopic hemogram. Recent combination of Raman spectroscopy with optical trapping techniques lay a corner stone for future automation of the spectral based cell differentiation10,12,35,36

In this study characteristic Raman spectral images from the major leukocyte subtypes (neutrophils and eosinophils as representatives of granulocytes and the mononuclear monocytes and lymphocytes) are presented. Furthermore, a robust statistical classification model based on Raman spectroscopic data is developed to differentiate between the two most abundant leukocytes in the blood: the neutrophils (ca. 50−70% of all leukocytes) and lymphocytes (25−30% of all leukocytes), with a special focus on CD4+ T-lymphocytes.



MATERIALS AND METHODS Cell Isolation and Preparation. Leukocytes have been isolated from peripheral blood of healthy volunteers (with Germanic and Dravidian ethnic background) with informed consent according to the Ethics Committee of the Jena University Hospital. An established protocol for leukocyte separation based on Ficoll density gradient centrifugation has been used.37−39 Briefly, 9 mL of blood in ethylenediaminetetraacetic acid (EDTA) was drawn freshly by venipuncture using the BD vacutainer system, diluted 1:1 with phosphate-buffered saline (PBS) and layered on an equal amount of Ficoll (LSM 1077, PAA, Germany) in a 50 mL tube. Centrifugation for 15 min at 1000g at 4 °C yielded four different layers which can be assigned from top to bottom to blood plasma, mononuclear cells (MNC), separation medium, and granulocyte pellets along with erythrocytes. MNCs were collected from the interphase between plasma and separation solution with a Pasteur pipet. The CD4+ T Cell Isolation Kit II (Miltenyi-Biotec, Bergisch Gladbach, Germany) was used to further isolate CD4+ Tlymphocytes from the MNCs using the magnetic cell sorting (MACS) technique. The precipitated granulocytes along with erythrocytes were collected in a separate falcon tube from which the latter ones were removed by NH4Cl cytolysis. Immediately after this centrifugation step the leukocytes were chemically fixed with 4% formaldehyde for 10 min to avoid activation of the cells during further sample treatment. This is especially important for the neutrophils as they are extremely sensitive to stress. Afterward, the cells were washed successively 5336

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342

Analytical Chemistry

Article

Figure 1. Morphological characteristics of the different leukocytes: neutrophil (A and E), eosinophil (B and F), monocyte (C and G), and lymphocyte (D and H). (A−D) white light images after Kimura staining, (E−H) false color Raman images of the same cell using the intensity at ∼788 cm−1 to highlight the nucleus (pink) and the intensity of the CH stretching between 2800 and 3050 cm−1 to color code the overall cell area (blue). (I) Averaged Raman spectra of the cytoplasm, nucleus, and background region.

pH 6.7, all from PAA Chemical) were placed on the cells and allowed to stand for 5 min. The stained cells were dip-washed gently with distilled water and allowed to dry at room temperature. The Raman mapped cells were relocated and the cell type assigned by investigation through a microscope (Axio Imager Z1, Carl Zeiss micro). Statistical Data Analysis. Data preprocessing and statistical analysis was carried out with Gnu R40 and the R package, ‘hyperSpec’.41 Unsupervised statistical classification methods, such as hierarchical cluster analysis (HCA) and principal component analysis (PCA) as well as the supervised statistical algorithm linear discriminant analysis (LDA) were applied to differentiate neutrophils and lymphocytes based on their Raman spectra. Raman maps have been recorded of 23 neutrophils and of 24 CD4+ T-lymphocytes, yielding in total 11621 Raman spectra of neutrophils and 3589 Raman spectra of the smaller T-lymphocytes. Raman spectra of the cells have been selected automatically from the Raman maps with a threshold criterion. High enough Raman intensity (threshold of 750 counts on the intensity scale bar) in the spectral region between 1100−1500 cm−1 could distinguish between cell and noncellular background region. For hierarchical cluster analysis (HCA) and principal component analysis (PCA) one averaged Raman spectrum from one whole cell area (about 100−500 individual spectra) was used, giving 23 average spectra of neutrophils and 24 average spectra of CD4+ T-lymphocytes. The background was corrected for each spectrum by fitting a polynomial baseline with 15 points included in the hyperSpec-package, and the spectra were vector normalized. The data were truncated keeping only the fingerprint region from 600 to 1800 cm−1. The spectral distances in the HCA were calculated based on the Euclidean distance, and Ward’s algorithm was used to build the cluster.

with PBS/2% FCS and 0.9% NaCl. For the Raman measurements 1 × 106 leukocytes (based on fixed cell count using a Neubauer chamber) were suspended in 100 μL of 0.9% NaCl and coated onto CaF2 slides by means of cytospin (Shandon Cytospin3 Cytocentrifuge, ThermoScientific, Waltham, USA, 6 min, 300 g). To ensure immobilization of the leukocytes the CaF2 slides have been precoated with 0.2% gelatin for 10 min. For recording high resolution Raman images of the leukocytes Ficoll separation was not performed. Erythrocytes have been removed from the EDTA blood by cytolysis with NH4Cl solution. After washing with PBS/2% FCS the leukocytes were resuspended in 0.9% NaCl and casted by means of cytospin onto gelatin-coated CaF2 slides. No chemical fixation was used. Raman spectroscopic analysis was completed within 4, maximal 6 h after blood withdrawal. The leukocyte subtypes have been assigned via Kimura staining after the Raman measurement. Raman Spectroscopy. Raman spectra of the cells coated on CaF2 slides were recorded with an upright micro-Raman setup (CRM 300, WITec GmbH, Germany) equipped with a 300 g/mm grating (7 cm−1 resolution) and a Deep Depletion CCD camera (DU401 BR-DD, ANDOR, 1024 × 127 pixels) cooled down to −75 °C. The cells were excited with a 785 nm diode laser which was focused through a Zeiss 50x objective (NA 0.95) onto the cells giving 75 mW in the object plane. Raman images for leukocyte differentiation were recorded in the scanning mode with a step size of 0.5 μm and integration time of 1 s per spectrum. For high-resolution Raman a Zeiss 100x objective (NA 0.9) and a step size of 0.3 μm was used. Identification of the Cell Type. The leukocyte subtype of the cells investigated by Raman spectroscopy was verified after the Raman measurements by Kimura staining. Ten microliters of the Kimura staining solution (toluidine blue, 0.03% light green SF yellowish, saturated saponin, and phosphate buffer, 5337

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342

Analytical Chemistry

Article

Figure 2. Dendrogram (A) and mean cluster spectra with standard deviation (B) of the hierarchical cluster analysis of neutrophils (red) and CD4+ T-lymphocytes (blue). Score plot (C) of the principal component analysis of the same cells. (D) Loadings of PC1 and PC3.



RESULTS Visualization of Leukocyte Morphology. The different subtypes of leukocytes are all distinct in form and function. In classical hematology differences in nucleus shape and the presence or absence of granules are visualized by staining methods, such as Kimura staining. Raman spectroscopy as a label-free but very sensitive method can not only visualize morphological features of cells without the need of staining but even the characteristic biochemical features. A comparison of bright field images of Kimura stains and false color high resolution Raman images of neutrophils, eosinophils, monocytes, and lymphocytes is shown in Figure 1. Due to their rareness (less than 1% of all leukocytes), basophils have been omitted from this study. In the bright field images (Figure 1A− D) the cells appear more or less round with an average diameter between 8 and 12 μm as is expected for leukocytes. Monocytes and lymphocytes are on average slightly smaller in size than the granulocytes. Kimura staining clearly highlights the nucleus in dark blue and helps to visualize the granules. In an unstained white light image those features would be hard to see. Raman spectra of the same cells had been recorded prior to Kimura staining. The integrated intensities of the CH stretching vibration between 2800 and 3050 cm−1 and of the nucleic acid vibration around 788 cm−1 were used to create the false color Raman maps shown in Figure 1E−H. Clearly, the boundary of the cells and the nucleus can be seen in these

A classification model was built for the two most abundant types of leukocytes in the blood using linear discriminate analysis (LDA). In total 15210 individual spectra of 47 maps (11621 spectra of 23 Raman maps from neutrophils and 3589 spectra from 24 Raman maps of CD4+ T-lymphocytes) were background corrected using the sensitive nonlinear iterative peak (SNIP) clipping algorithm42 and vector normalized. To reduce the dimension of the data prior to LDA, PCA analysis was carried out, and the first 12 PCs were used to build the LDA classification model. The model was cross-validated using the leave-one-out cross-validation method leaving out all spectra of an entire cell. The above-described PCA-LDA was tested with a completely independent test data set using leukocytes isolated on a different day from the blood of two different healthy persons, one of which has not contributed to the training data set at all. Furthermore, the two healthy volunteers are representatives of two different ethnic groups (Germanic and Dravidian). The sample preparation was the same as described in the sample preparation section for the neutrophils, and the spectral preprocessing was done in a similar manner as described above. Raman spectra from 37 Raman maps containing a mixed population of neutrophils and lymphocytes were projected on the PCA build with the training set, and the corresponding scores were fed into the model. The cell type assigned by the algorithm was subsequently confirmed by the standard Kimura staining method. 5338

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342

Analytical Chemistry

Article

statistical method was used. Based on the spectral differences and similarities present in the Raman spectra, two well separated clusters have been formed as shown in the dendrogram in Figure 2A. Similar spectra are clustered together, while spectra with a high dissimilarity are further away in the cluster. A very good separation without any misclassification between the neutrophils and the CD4+ Tlymphocytes is achieved. Figure 2B shows the averaged spectra from each cluster together with their respective standard deviation. As there are no misclassifications in the HCA, those cluster spectra represent the average spectra of the two cell populations. Raman bands representing different cellular components are marked in the figure. As discussed for Figure 1 spectral contribution from the cellular components are visible as expected for a human eukaryotic cell. Typical vibrational contributions due to proteins are the amide I band (around 1660 cm−1), the amide II band (around 1257 cm−1), and contributions from individual amino acids, such as phenylalanine (1002 cm−1) and tyrosine (1605 cm−1). As those spectra are average spectra from the whole cell also nucleic acid contributions are visible, such as ring breathing modes of the nucleic acid bases (678 cm−1 guanine ring breathing, 730 cm−1 adenine ring breathing,17,43 and at 1576 cm−1 guanine and adenine contributions17,43) and from the O−P−O backbone (788 cm−117,43 and 1095 cm−117,43). The intense vibrational band around 1341 cm−1 has contributions from proteins and nucleic acids and to a minor extent also from carbohydrates. One of the major differences between neutrophils and lymphocytes are the intensities of the vibrational bands due to nucleic acids. Those bands are much more prominent in the spectra of the CD4+ T-lymphocytes than in the neutrophils. This can be explained by a much larger nucleus in the lymphocytes which almost fills the whole cell compared to a smaller lobed nucleus in the neutrophils. Contributions from protein components are more pronounced in the neutrophil spectra. The tyrosine vibrations at 855 cm−1 and around 1605 cm−1 can be clearly identified in neutrophil spectra compared to the CD4+ T-lymphocytes spectra. Carotenoid bands (1521 and 1164 cm−1) are present in some CD4+ T-lymphocytes with different concentrations but are not found in the neutrophils. Not all CD4+ T-lymphocytes contain carotinoids, which results in a high standard deviation in the average spectra for these two vibrational bands. This is in good agreement with previous investigations of lymphocytes.19,21 Other differences between the two cell types are found around 663 cm−1 (tentatively assigned to the cystine vibrations) and in the spectral fine structure between 800 and 900 cm−1 and between 1500 and 1600 cm−1. Principal Component Analysis. The good classification results of the HCA can be reproduced with principal component analysis (PCA) as well using the same averaged preprocessed Raman spectra of the neutrophils and the CD4+ T-lymphocytes. A very good separation of the two leukocyte subgroups can be achieved along the first principal component (PC1) which describes about 74% of the spectral variance (Figure 2C). The higher PCs contain mainly intraclass variations. The second principal component (PC2) captures ∼11%, and PC3 contains ∼6% of the spectral variances in the Raman data. The two-dimensional score plot of PC 1 against PC 3 is shown in Figure 2C; the corresponding loadings of PC1 and PC3 are shown in Figure 2D. In the loadings of PC1, which can discriminate between neutrophils and CD4+ T-lymphocytes, the same spectral features are identified as were found to

pictures. Neutrophils are characterized by a multilobed nucleus, such as the three-lobed nucleus in Figure 1A and 1E, and eosinophils have a bilobed nucleus (Figure 1B and F). Monocytes and lymphocytes are mononuclear cells, i.e. their nucleus is only single-lobed. While monocytes have a kidney shaped nucleus (Figure 1C and G), lymphocytes have a round, askew nucleus which fills almost the whole cell (Figure 1D and H). A very good agreement is found between the false color Raman maps and the Kimura stained images which are traditionally used to assign unambiguously the correct leukocyte subtype. This highlights the power of Raman spectroscopy for label-free cell characterization. From the Raman data not only morphological information but also information on the chemical composition of each cell can be obtained. The averaged Raman spectra in the fingerprint region (600−1800 cm−1) from the different cellular regions are depicted exemplarily for each cell type in Figure 1I. The Raman signal of the CaF2 substrate precoated with gelatin gives nearly constant background signal. The sharp CaF2 Raman band of the substrate at 322 cm−1 does not lie within the Raman spectral region (600−1800 cm−1) chosen for statistical analysis but can be used as an internal wavenumber standard. The overall features of the Raman spectra from the nuclear region are similar for all leukocytes as is expected from the similar chemical composition made of polynucleotides and proteins. Characteristic vibrational bands that can be assigned to polynucleotides are found at 730 cm−1 (adenine ring breathing17,43), around 788 cm−1 (contributions from O−P− O backbone ring and nucleic acid base vibrations17,43), around 1097 cm−1 (backbone and symmetric O−P−O vibration17,43), and around 1576 cm−1 (guanine and adenine contributions17,43). The spectral differences between the Raman spectra of the different nuclei of the leukocyte subtypes are subtle and are not easily spotted by eye. In the Raman spectra originating from the cytoplasmic region (top spectra in Figure 1I) slight variations in peak positions and relative intensities are observed. In all spectra typical Raman bands from proteins, lipids, and carbohydrates can be seen. Very prominent in all spectra are the amide I band around 1663 cm−1, the CH deformation band around 1450 cm−1, and the phenylalanine band around 1002 cm−1. The most obvious spectral difference can be seen in the averaged Raman spectrum of the lymphocyte which shows sharp Raman bands at 1522 cm−1 and 1158 cm−1 which are not observed in the Raman spectra of the other cell types examined. These features can be assigned to carotenoids. The presence of carotenoids in socalled Gall bodies in varying concentrations in some lymphocytes has been observed before.17,19,21 Other spectral regions which show significant variations in band position and relative intensities between the leukocytes subtypes are e.g. around 830−850 cm−1 and between 1520−1620 cm−1. These differences in the Raman spectra of the leukocytes originate from a different chemical composition (e.g., enzymes in the granules33,44) and point to the potential of Raman spectroscopy to differentiate various cell types. For the two most abundant leukocytes, the neutrophils and the lymphocytes, a classification model was built which can automatically assign the cell type from a mixed population of neutrophils and lymphocytes based on the Raman spectra. Details are explained in the next section Raman Spectroscopic Differentiation of Neutrophils and Lymphocytes. Hierarchical Cluster Analysis. To visualize spectral differences and cluster formation according to the cell type hierarchical cluster analysis (HCA), an unsupervised 5339

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342

Analytical Chemistry

Article

Table 1. Confusion Table for the Classification (Left Columns) and Identification (Right Columns) Using Linear Discriminate Analysisa true classification

identification

single spectra

neutrophils

CD4+ T-lymphocytes

neutrophils

lymphocytes

Predicted neutrophils CD4+ T-lymphocytes

11147 474

353 3236

4486 1291

1339 6914

true classification majority vote for each cell Predicted neutrophils CD4+ T-lymphocytes a

identification +

neutrophils

CD4 T-lymphocytes

neutrophils

lymphocytes

23 0

0 24

11 1

0 25

The two top four-field-tables give the results based on individual spectra; the two bottom tables are based on the majority vote for each cell.

be the characteristic differences between the averaged cluster spectra from the hierarchical cluster analysis (Figure 2B). The loadings of PC1 show intense positive features at 789 cm−1 and 1092 cm−1. This indicates higher nucleic acid contributions in the cells toward positive PC1, which are the lymphocytes. Other positive features can be found in the region of the amide I band centered around 1677 cm−1. This could be a hint on different protein expression profiles with different secondary structure in the lymphocytes. The loadings of PC3 are dominated by the vibrational features of carotenoids with negative bands around 1523 and 1158 cm−1. Especially the two CD4+ T-lymphocytes in the second quadrant of the score plot (Figure 2C) are rich in carotenoids and well separated from the other CD4+ T-lymphocytes along PC3. Classification Model Based on Linear Discriminant Analysis. The previously applied statistical methods, HCA and PCA, are unsupervised statistical methods which are used to visualize the data set and characterize the classification behavior. However, those methods are not suited to build classification models for the prediction of the identity of unknown cells. Therefore, linear discriminant analysis (LDA), a supervised statistical method, has been employed to develop a classification model for the differentiation of neutrophils and lymphocytes. As training data set the previously shown data were used. However, not the averaged spectra were used but each individual cell spectrum from the Raman map. The classification model was built with 15210 spectra and cross-validated by leaving out each time the spectra of one entire scan. The results are shown in Table 1 (top left) and visualized in the score plot of the classification model in Figure 3. A clear separation of neutrophils and CD4+ T-lymphocytes with only a few misclassifications can be seen. Out of totally 11621 neutrophil spectra 11147 spectra (95.9%) were assigned correctly, and only 474 Raman spectra (4.1%) were misclassified and assigned to the T-lymphocytes. For the CD4+ T-lymphocytes, 3236 spectra (90.2%) out of the totally 3589 lymphocyte spectra have been assigned correctly, and 353 Raman spectra of CD4+ T-lymphocytes (9.8%) were wrongly classified as neutrophils. This yields an accuracy of 94.5% of this classification model (Table 1, top left). This is a very good value, especially when taking into account single spectra recorded from the cells were used for classification. This can find application in the design of future experiments when the goal is to increase the throughput and reduce the number of

Figure 3. LDA scatter plot of neutrophils (red circles) and CD4+ Tlymphocytes (green triangles).

spectra that needs to be measured from one cell for unambiguous identification of the leukocyte subtype. The classification accuracy can be increased to 100% if the majority vote is used for classification of an entire cell into the respective leukocyte subgroup as can be seen in the confusion Table 1 (bottom left). The classification model was tested with an entirely new and independent data set consisting of 37 cells, which were recorded on a different day. The cells were a mixed population of neutrophils and lymphocytes isolated from two healthy volunteers, of which one has not contributed with his blood to the test data set and belongs to a different ethnic group. The membership of the leukocyte to a certain subpopulation was assigned after the Raman measurement by staining the cells with Kimura’s staining solution and observing the nuclear morphology of the stained cell. Out of 37 cells, 12 cells were identified as neutrophils and 25 cells were identified as lymphocytes. The Raman data were treated in the same way as it was done for the classification data set, and identification was carried out by using all 14030 individual spectra belonging to the cellular region of the Raman maps. The results of the 5340

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342

Analytical Chemistry

Article

characterized diseases or syndromes have to be included in the statistical model. It is imaginable that the specific leukocyte response to a certain challenge, e.g. binding of ‘pathogen associated molecular patterns’, will be revealed in the Raman spectra and can provide additional information on different diseases and syndromes to the physicians immediately after recording such a spectroscopic hemogram.

identification are presented in Table 1 (top right). Out of 5777 total neutrophil spectra 4486 spectra (77.6%) have been classified correctly, and out of 8253 lymphocyte spectra 6914 spectra (83.8%) have been assigned correctly, giving an overall accuracy of the classification model of 81.2%. The reduced identification accuracy compared to the training data set (Table 1, top left) are mainly attributed to the biological variance (different blood donors on different days) and the daily variations in the sample preparation. Variations due to instrument variability are considered to be only minute. The classification accuracy can be improved if not only a single spectrum, but the majority vote of all the spectra from one cell is used to predict the leukocyte cell type for this entire cell. The results are shown in the confusion Table 1 (bottom right). Only one neutrophil has been misclassified as lymphocyte giving an overall accuracy of 97.3%. This indicates the spectral variance due to biological variance (different donors from different ethnic groups); minute changes in the sample preparation method as well as day-today changes in the spectrometer performance are smaller than the characteristic spectral differences between lymphocytes and neutrophils. Also, the identification of the lymphocytes with high accuracy based on the classification done using one of its subtype, CD4+ T-lymphocytes indicates that the lymphocytes are not only morphologically but chemically very different from neutrophils. This change in the biochemical components is very well brought out by Raman spectroscopy. This is a very promising result for future spectroscopic assignment of the leukocyte subpopulation without the need of any staining solution.



AUTHOR INFORMATION

Corresponding Author

*Phone: +49-3641-9323364. Fax: +49-3641-9323382. E-mail: [email protected]. Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS Financial support of the BMBF via the integrated research and treatment center “Center for Sepsis Control and Care” (FKZ 01EO1002) and from the EU within the “Framework Programme 7” (FP7, P4L grant agreement no.: 224014) is highly acknowledged. The authors are indebted to Yvonne Schlenker for her excellent technical assistance, to Claudia Beleites for helping with the statistical programming, and to Christian Matthäus for his valuable advice in cell handling. A.R. and U.N. contributed equally to this work. M.B. and J.P. share the senior authorship.



REFERENCES

(1) Borregaard, N. Eur. J. Haematol. 1988, 41 (5), 401−413. (2) Brakhage, A. A.; Bruns, S.; Thywissen, A.; Zipfel, P. F.; Behnsen, J. Curr. Opin. Microbiol. 2010, 13 (4), 409−415. (3) Remijsen, Q.; Kuijpers, T. W.; Wirawan, E.; Lippens, S.; Vandenabeele, P.; Berghe, T. V. Cell Death Differ. 2012, 18 (4), 581− 588. (4) Papayannopoulos, V.; Zychlinsky, A. Trends Immunol. 2009, 30 (11), 513−521. (5) Brinkmann, V.; Zychlinsky, A. Nat. Rev. Microbiol. 2007, 5 (8), 577−582. (6) Mantovani, A.; Cassatella, M. A.; Costantini, C.; Jaillon, S. Nat. Rev. Immunol. 2011, 11 (8), 519−531. (7) Eliasson, C.; Engelbrektsson, J.; Lorén, A.; Abrahamsson, J.; Abrahamsson, K.; Josefson, M. Chemom. Intell. Lab. Syst. 2006, 81 (1), 13−20. (8) Brown, K. L.; Palyvoda, O. Y.; Thakur, J. S.; Nehlsen-Cannarella, S. L.; Fagoaga, O. R.; Gruber, S. A.; Auner, G. W. J. Immunol. Methods 2009, 340 (1), 48−54. (9) Brown, K. L.; Palyvoda, O. Y.; Thakur, J. S.; Nehlsen-Cannarella, S. L.; Fagoaga, O. R.; Gruber, S. A.; Auner, G. W. Cytometry, Part A 2009, 75A (11), 917−923. (10) Aseefhali Bankapur, E. Z.; Chidangil, S.; Valiathan, M.; Mathur, D. PloS One 2010, 5 (4), e10427. (11) Eliasson, C.; Lorén, A.; Engelbrektsson, J.; Josefson, M.; Abrahamsson, J.; Abrahamsson, K. Spectrochim. Acta, Part A 2005, 61 (4), 755−760. (12) Samantha Fore, J. C.; Taylor, D.; Huser, T. J. Opt. (Bristol, U. K.) 2011, 13 (4), 044021. (13) Harz, M.; Kiehntopf, M.; Stockel, S.; Rosch, P.; Deufel, T.; Popp, J. Analyst 2008, 133 (10), 1416−1423. (14) Neugebauer, U.; Bocklitz, T.; Clement, J. H.; Krafft, C.; Popp, J. Analyst 2010, 135 (12), 3178−3182. (15) Neugebauer, U.; Clement, J. H.; Bocklitz, T.; Krafft, C.; Popp, J. J. Biophotonics 2010, 3 (8−9), 579−587. (16) Otto, C.; Sijtsema, N. M.; Greve, J. Eur. Biophys. J. 1998, 27 (6), 582−589. (17) Pully, V. V.; Lenferink, A. T. M.; Otto, C. J. Raman Spectrosc. 2011, 42 (2), 167−173.



CONCLUSION AND OUTLOOK It could be shown that Raman spectroscopy has been able to identify and distinguish leukocytes from different subgroups. False color Raman images could reveal the same morphological features of the leukocytes as usually visualized with classical staining methods, however, without the need of any additional labels. With the aid of statistical models it was possible to discriminate between different cell types based on the specific biochemical information present in the Raman spectra. In this proof-of-principal study unsupervised and supervised statistical methods yielded a good differentiation of the two most abundant leukocyte subtypes, the neutrophils and lymphocytes based on the Raman data. PCA and HCA could extract specific chemical features, such as the presence of carotenoids in some lymphocytes from the Raman data. A robust classification model could be built with PCA-LDA. This model was tested with a completely independent test data set using leukocytes isolated on a different day from the blood of two different healthy persons, one of which has not contributed to the training data set at all and belongs to a different ethnic group. The high identification accuracy (97.3%) indicates that the heterogeneity due to person-to-person variability, different ethnic groups (Germanic and Dravidian), and sample preparation on different days has minimum influence. Nevertheless, in order to assess the span of the biological heterogeneity due to different age, sex, life style, and ethnic background the current data set will be systematically increased. This study is highly promising and shows the high capability of Raman spectroscopy for the differentiation of leukocyte subpopulations without the need of external markers. This paves the way for automated analysis toward a spectroscopic hemogram where also leukocytes from patients with well5341

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342

Analytical Chemistry

Article

(18) Puppels, G. J.; Garritsen, H. S.; Segers-Nolten, G. M.; de Mul, F. F.; Greve, J. Biophys. J. 1991, 60 (5), 1046−1056. (19) Puppels, G. J.; Garritsen, H. S. P.; Kummer, J. A.; Greve, J. Cytometry 1993, 14 (3), 251−256. (20) Ramanauskaite, R. B.; Segers-Nolten, I. G. M. J.; de Grauw, K. J.; Sijtsema, N. M.; van der Maas, L.; Greve, J.; Otto, C.; Figdor, C. G. Pure Appl. Chem. 1997, 69 (10), 2131−2134. (21) Schut, T. C. B.; Puppels, G. J.; Kraan, Y. M.; Greve, J.; van der Maas, L. L. J.; Figdor, C. G. Int. J. Cancer 1997, 74 (1), 20−25. (22) Sijtsema, N. M.; Tibbe, A. G. J.; Segers-Nolten, I. G. M. J.; Verhoeven, A. J.; Weening, R. S.; Greve, J.; Otto, C. Biophys. J. 2000, 78 (5), 2606−2613. (23) Smith, Z. J.; E. W., J.-C.; Quataert, S. A.; Berger, A. J. J. Biomed. Opt. 2010, 15 (3), 036021. (24) Takai, Y.; Masuko, T.; Takeuchi, H. Biochim. Biophys. Acta, Gen. Subj. 1997, 1335 (1−2), 199−208. (25) van Manen, H.-J.; Kraan, Y. M.; Roos, D.; Otto, C. Proc. Natl. Acad. Sci. U.S.A. 2005, 102 (29), 10159−10164. (26) Mostaco-Guidolin, L. B.; Bachmann, L. Appl. Spectrosc. Rev. 2011, 46 (5), 388−404. (27) Liu, K. Z.; Xu, M. Q.; Scott, D. A. Br. J. Hamaetol. 2007, 136 (5), 713−722. (28) Liu, K. Z.; Shi, M. H.; Mantsch, H. H. Blood Cells, Mol., Dis. 2005, 35 (3), 404−412. (29) Babrah, J.; McCarthy, K.; Lush, R. J.; Rye, A. D.; Bessant, C.; Stone, N. Analyst 2009, 134 (4), 763−768. (30) Vahlsing, T.; Damm, U.; Kondepati, V. R.; Leonhardt, S.; Brendel, M. D.; Wood, B. R.; Heise, H. M. J. Biophotonics 2010, 3 (8− 9), 567−578. (31) Mannie, M. D.; McConnell, T. J.; Xie, C.; Li, Y.-q. J. Immunol. Methods 2005, 297 (1−2), 53−60. (32) van Manen, H.-J.; Uzunbajakava, N.; van Bruggen, R.; Roos, D.; Otto, C. J. Am. Chem. Soc. 2003, 125 (40), 12112−12113. (33) Sijtsema, N. M.; Otto, C.; Segers-Nolten, G. M. J.; Verhoeven, A. J.; Greve, J. Biophys. J. 1998, 74 (6), 3250−3255. (34) Zoladek, A. B.; Johal, R. K.; Garcia-Nieto, S.; Pascut, F.; Shakesheff, K. M.; Ghaemmaghami, A. M.; Notingher, I. Analyst 2010, 135 (12), 3205−3212. (35) Chan, J. W.; Taylor, D. S.; Lane, S. M.; Zwerdling, T.; Tuscano, J.; Huser, T. Anal. Chem. 2008, 80 (6), 2180−2187. (36) Dochow, S.; Krafft, C.; Neugebauer, U.; Bocklitz, T.; Henkel, T.; Mayer, G.; Albert, J.; Popp, J. Lab Chip 2011, 11 (8), 1484−1490. (37) Workalemahu, G.; Foerster, M.; Kroegel, C. J. Cell. Physiol. 2006, 207 (1), 67−74. (38) Workalemahu, G.; Foerster, M.; Kroegel, C.; Braun, R. K. J. Immunol. 2003, 170 (1), 153−157. (39) Workalemahu, G.; Foerster, M.; Kroegel, C. J. Leukocyte Biol. 2004, 75 (4), 657−663. (40) R Development Core Team, R. R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing: Vienna, Austria, 2011. (41) Beleites, C.; Sergo, V. Journal of Statistical Software 2011. (42) Ryan, C. G.; Clayton, E.; Griffin, W. L.; Sie, S. H.; Cousens, D. R. Nucl. Instrum. Methods Phys. Res., Sect. B 1988, 34 (3), 396−402. (43) Uzunbajakava, N.; Lenferink, A.; Kraan, Y.; Willekens, B.; Vrensen, G.; Greve, J.; Otto, C. Biopolymers 2003, 72 (1), 1−9. (44) van Manen, H.-J.; Kraan, Y. M.; Roos, D.; Otto, C. J. Phys. Chem. B 2004, 108 (48), 18762−18771.

5342

dx.doi.org/10.1021/ac3007363 | Anal. Chem. 2012, 84, 5335−5342