MALDI-imaging segmentation is a powerful tool for spatial functional proteomic analysis of human larynx carcinoma

Abstract

Purpose

For several decades, conventional histological staining and immunohistochemistry (IHC) have been the main tools to visualize and understand tissue morphology and structure. IHC visualizes the spatial distribution of individual protein species directly in tissue. However, a specific antibody is required for each protein, and multiplexing capabilities are extremely limited, rarely visualizing more than two proteins simultaneously. With the recent emergence of matrix-assisted laser desorption/ionization imaging mass spectrometry (MALDI-imaging), it is becoming possible to study more complex proteomic patterns directly in tissue. However, the analysis and interpretation of large and complex MALDI-imaging data requires advanced computational methods. In this paper, we show how the recently introduced method of spatial segmentation can be applied to analysis and interpretation of a larynx carcinoma section and compare the spatial segmentation with the histological annotation of the same tissue section.

Methods

Matrix-assisted laser desorption/ionization imaging is a label-free spatially resolved analytical technique, which allows detection and visualization of hundreds of proteins at once. Spatial segmentation of the MALDI-imaging data by clustering of spectra by their similarity was performed, automatically generating spatial a segmentation map of the tissue section, where regions of similar proteomic patterns were highlighted. The tissue was stained with the hematoxylin and eosin (H&E), histopathologically analyzed and annotated. The segmentation map was interpreted after its overlay with the H&E microscopy image.

Results

The automatically generated segmentation map exhibits high correspondence to the detailed histological annotation of the larynx carcinoma tissue section. By superimposing, the segmentation map based on the proteomic profiles with H&E-stained microscopic images, we demonstrate precise localization of complex and histopathologically relevant tissue features in an automated way.

Conclusions

The combination of MALDI-imaging and automatic spatial segmentation is a useful approach in analyzing carcinoma tissue and provides a deeper insight into the functional proteomic organization of the respective tissue.

Introduction

For a long time, the evaluation of tissue samples has been limited to histomorphological methods. Fresh frozen or chemically fixed tissue are sectioned, stained, and subsequently examined for structural, anatomical or histological features, and pathological changes of those. In a clinical setting, these changes are graded according to commonly accepted guidelines, which in most cases represent the main method for the assessment of tumors or tumor-suspicious tissue regions and are the basis of diagnostically or therapeutically relevant decisions. The localization of specific proteins in tissue on a cellular and subcellular spatial resolution is possible by immunohistochemistry (IHC). However, IHC is a targeted method, which means the identity of the proteins of interest must be known and protein-specific antibodies are necessary. IHC is not suited for discovery (i.e., untargeted) studies, in which new biomarkers, for example, for cancer diagnosis are sought for. Moreover, IHC is limited to visualizing a few proteins in parallel and therefore does not provide a global proteomic view of the tissue.

Molecular genomic and proteomic-based attempts that try to refine or facilitate these histopathological analyses have been successful only to minor degree (Rezende et al. 2010). Large-scale differential studies, both on the transcript (e.g., microarrays) and protein level (e.g., 2D gel or chromatography-based proteomics) are generally prone to oversimplification, as the complex morphology of tissue biopsies are not accounted for when selecting the samples. As a consequence, the expression profiles generated typically represent an average across multiple morphological features and cell types. This limitation can be partially overcome by tissue microdissection or laser-capture microdissection, which allows a more precise sample preparation for further molecular analysis (von Eggeling et al. 2007). However, tissue microdissection is a time-consuming technique requiring pathological competence and, because of limited sample amounts, highly sensitive downstream analysis. Sensitivity of genomic and proteomic technologies is continuously increasing, especially mass spectrometric techniques, like surface-enhanced laser desorption/ionization time of flight mass spectrometry (SELDI-TOF-MS) using affinity chromatographic surfaces, which are commonly used to investigate differentially expressed proteins in body fluids, cells, and tissue (Paradis et al. 2005; Escher et al. 2006; Ward et al. 2006; von Eggeling et al. 2007). About 3000 cells are needed to receive an adequate proteomic profile (Opitz et al. 2004).

Mass spectrometry-based imaging techniques capable of elucidating molecular profiles directly from tissue are becoming increasingly popular. Matrix-assisted laser desorption/ionization (MALDI) imaging mass spectrometry (IMS), shortly called as MALDI-imaging, is the most popular IMS technique for proteomic applications and allows to obtain proteomic profiles directly from thin tissue sections (Chaurand et al. 2004). MALDI-imaging is a label-free technique and can visualize the distribution of hundreds of molecular compounds in a single measurement, while maintaining the morphological and molecular integrity of the tissue. Mass spectra are acquired across a flat sample (e.g., a tissue section); each mass spectrum is measured at a pixel with assigned pair of spatial coordinates x and y. Lateral resolution (size of a pixel) is currently as low as 20 μm for the analysis of proteins (Lagarrigue et al. 2011), and 5 μm for phospholipids and neuropeptides (Rompp et al. 2010). For a fixed molecular mass (mass over charge value, or m/z-value, is usually interpreted in MALDI as the molecular mass since ions with a charge of +1 prevail), spectral intensity is correlated with the concentration of a molecular compound. Images are typically generated by visualizing the intensity of a selected mass spectrometric peak as a color gradient across the analyzed area. Several of these so-called m/z-images can be easily overlaid, creating multiplexed images simultaneously visualizing the distribution of any combination of compounds. In addition, m/z-images can easily be superimposed with high-resolution optical microscopy image of the same section, which is stained histologically after MALDI acquisition allowing precise matching between molecular localization and morphological features. We and others have used MALDI-imaging for the pathological analysis of various tissue types (Deininger et al. 2008; Walch et al. 2008; Chaurand et al. 2004; Rauser et al. 2010; Cazares et al. 2011).

A typical MALDI-imaging data set comprises several thousand individual mass spectra, with several thousand intensity values each, and thus has several million individual data points. Although the superimposition of individual m/z-images with microscopic images can provide valuable insight, efficient mining of a large MALDI-imaging data set cannot be performed manually, especially if multiple measurements are involved. Moreover, the common approach to select m/z-values of interest based on visual examination of the average spectrum can cause a compound present at a small portion of data points to be overlooked, because its average intensity is low as compared to a baseline peak present at data points. One of the computational methods proposed for MALDI-imaging data mining is the spatial segmentation. The individual spectra of a data set are grouped (e.g., by a clustering method) according to their spectral similarity, and groups of similar spectra are highlighted on the tissue in a common color (Walch et al. 2008; Deininger et al. 2008). As physiological similarity in the tissue is usually reflected in mass spectral similarity, spatial segmentation is a straightforward approach to detect tissue features in an automated way. However, MALDI-imaging data exhibit strong pixel-to-pixel variation, which complicates both visual examination and computational analysis. We have recently proposed a computational method significantly reducing the pixel-to-pixel variation, leading to clear, noiseless, smooth segmentation maps revealing a complex and detailed histological structure (Alexandrov et al. 2010) with a proof-of-principle application in histology. Later, we proposed improved peak picking for MALDI-imaging data, which leads to more sensitive results (Alexandrov and Kobarg 2011). In this paper, we apply the spatial segmentation procedure from Alexandrov et al. (2010) with improved peak picking to a complex structured larynx carcinoma sample, supported by examination and interpretation of the histological and MALDI-imaging data by an experienced pathologist (GE). Whereas the publication (Alexandrov et al. 2010) was the proof-of-principle application of segmentation to MALDI-imaging data with only limited histological insight (tumor tissue was discriminated from nontumor), this paper explores and demonstrates the potential of the segmentation approach to histological evaluation. Moreover, the segmentation maps demonstrate the correlation of proteomic images and detailed tissue topology. A detailed localization of functional proteomic domains through using MALDI-imaging data segmentation is important for new diagnostic strategies and for biological understanding of tumors.

Materials and methods

Samples collection and histological annotation

Tissue samples from patients with larynx carcinoma were obtained from the Department of Otorhinolaryngology of the University Hospital Jena. All necessary approval was obtained from the local Ethics Committee, approval No. 3008-12/10. After surgical resection, the samples were snap frozen in liquid nitrogen and stored at −80 °C until analysis. Tumor specimens were categorized according to the WHO classification (Sobin and Wittekind 2009). A detailed histological analysis of the cryosections was performed by an experienced pathologist (GE) after postacquisition hematoxylin and eosin (H&E) staining. Functional areas like epithelium or stroma were marked in digital images. After histological analysis, the most representative section showing at the same time the sufficient level of histological complexity was selected, and a consecutive section was made for MALDI-imaging analysis. This sample was obtained from a male, 56-year-old patient and categorized according to the AJCC classification from 2002 (sixth edition) as squamous cell carcinoma of the larynx. The tumor was classified as TNM pT3pN2bM0 R0 L1, UICC stage IV, and grading G2.

MALDI-imaging analysis

For an overview of the MALDI-imaging analysis, see Fig. 1. The cryosection of 10 μm was prepared and thaw-mounted on precooled conductive glass slide (Bruker Daltonik GmbH, Germany). To prevent samples deterioration, we stored the tissue specimen at −80 °C and subjected it to −20 °C only in the cryotome, which is the standard protocol for sample preparation for MALDI-imaging. The section was briefly washed twice for 1 min each in 70 % ethanol, and once for 1 min in absolute ethanol, then dried in a vacuum desiccator. The matrix (10 mg/mL sinapinic acid in 60 acetonitrile and 40 % water with 0.2 % trifluoroacetic acid) was applied using the ImagePrep device (Bruker Daltonik) following the standard protocol. Mass spectra were acquired using a MALDI-TOF mass spectrometer autoflex speed LRF (Bruker Daltonik) equipped with a 1 kHz smartbeam II laser. The instrument was operated in the linear mode and positive polarity. Spectra were acquired in the mass range m/z 2,000–20,000 at the sampling rate of 0.13 GS/s, 8,192 data points were acquired per spectrum. The MALDI-imaging raster size was set to 50 μm. 300 laser shots were summarized per raster spot. Acquisition and visualization was carried out using the flex Imaging 3.0 software (Bruker Daltonik). After MALDI-imaging analysis, the section was H&E-stained, scanned using a digital MIRAX scanner (Carl Zeiss, Goettingen, Germany) and histologically analyzed with the MIRAX software (Carl Zeiss) by an experienced histologist. The high-resolution H&E image and histological annotation were overlaid with MALDI-imaging data using the flex Imaging software (Bruker Daltonik).

Fig. 1.

MALDI-imaging data acquisition workflow and data representation as a data cube or a hyperspectral image with spatial coordinates x and y and with the mass spectral coordinate m/z. The brain section is shown for illustration only. In the current study, a human larynx carcinoma tissue section was analyzed instead (color figure online)

Mass spectrometry data preprocessing

Spectra were converted into mzXML format using the CompassXport software (Bruker Daltonik); all other processing was done using custom made scripts in Matlab (The Mathworks Inc., Natick, MA, USA). Spectra were loaded using the mzxmlread Matlab routine (Bioinformatics Toolbox) and interpolated to 10,000 bins, which is the standard number of bins used in our segmentation pipeline. After normalization to the total ion count (TIC), each spectrum was baseline corrected with the Matlab routine “msbackadj” (Bioinformatics Toolbox) with the “pchip” regression method, “lowess” smoothing method, and window size of 200 m/z. For each data set, peak picking was performed by selecting data set relevant peaks using the orthogonal matching pursuit method with subsequent selection of consensus peaks and their alignment to the peaks of the data set mean spectrum, as described in Alexandrov and Kobarg (2011). For each 10th spectrum (in chronological order), we selected its 100 most prominent peaks based on the peak shape and intensity; see Alexandrov et al. (2010) for a discussion on the peak picking procedure. As consensus peaks, we have selected those detected in at least 1 % of the considered spectra. As a peak width normally exceeds a m/z-bin size, this approach can select several m/z-values per peak and thereby exaggerate the influence of large and broad peaks in the subsequent statistical analysis. Alignment of the selected m/z-values to the peaks of mean spectrum and selection of only one m/z-value per peak, as introduced in Alexandrov and Kobarg (2011), solves this problem. Moreover, as these consensus peaks are selected using a threshold, omission of redundant peaks makes the overall peak picking more sensitive.

Spatial segmentation of MALDI-imaging data

For each data set, we applied the spatial segmentation method as proposed in (Alexandrov et al. 2010). Shortly, we performed peak picking, then applied edge-preserving denoising to the m/z-images of detected peaks, then clustered the reduced and processed spectra by their similarity. Moderate edge-preserving image denoising was applied to the data prior to clustering. Instead of High-dimensional Discriminant Data Clustering proposed (Alexandrov et al. 2010), we have used the k-means clustering for faster processing. In Alexandrov et al. (2010), k-means delivered poor results. In the current study, the segmentation maps produced with k-means, and HDDC are comparable (results not shown). We hypothesize that improvement of k-means segmentation maps became possible due to peak alignment, which reduces the influence of large peaks because only one aligned m/z-value was selected per peak. The k-means clustering was conducted with 10 replicates, “cluster” initialization, and “city block” distance; see Matlab documentation for information on these parameters. The segmentation results are presented as a segmentation map, with pseudo-colors assigned to pixels. We did not use any computational method for selection of the number of clusters automatically, but generated segmentation maps for 2–10 clusters, which were subsequently analyzed visually by an experienced pathologist (GE). After segmentation, we determined individual m/z-values co-localized with the segments of the spatial map corresponding to each cluster, by calculating the Pearson correlation between the cluster spatial mask (an image = 1 at the pixels of the cluster and 0 otherwise) and each m/z-image. Significantly correlated (p < 0.05) m/z-values with high correlation coefficients were selected and aligned to the peaks of the mean spectrum as done after peak picking (see above). The peak alignment reduces the number of co-localized m/z-values by omitting redundant m/z-values and selecting only one m/z-value per peak thus simplifying interpretation.

Results

A complex larynx carcinoma specimen containing normal squamous epithelium, different stages of dysplasia, invasive well differentiated and dissociated growing tumor parts with highly irregular cell and tissue architecture and surrounding stroma was used as an example to demonstrate the advantages of segmentation maps (Fig. 2a). A 10-μm-thick cryosection of this specimen was analyzed by MALDI-imaging. The resulting spectra were subjected to spatial segmentation. Ten clusters were allocated to the section (Fig. 2b) and visualized in the imaging software. Cluster 1 (brown in Fig. 2b) displays peripheral small infiltrates of dissociated growing cancer components and central preexistent connective tissue. Cluster 2 (dark red in Fig. 2b) is composed of preexistent connective tissue. The composition of the clusters 3–8 is described in detail in Fig. 3. It is of special interest that distinct histopathological differentiations of the epithelium (dysplasia, intraepithelial carcinoma) are clearly represented in corresponding map segments (clusters 6, 7, 8). The dysplastic epithelium and parts of the stroma adjacent to dysplastic epithelium are located in cluster 7 (Fig. 3b and c). Moreover, cluster 7 contains stroma of dissociated growing tumor components and preexistent connective tissue. The peripheral stroma is intermingled with dense tumor cell complexes. On the other hand, epithelial components and stroma are not separated in the still differentiated tumor component (cluster 8). Cluster 8 apparently comprises m/z-signals characterizing the still differentiated components of the carcinoma. Epithelial differentiations like epithelial dysplasia and intraepithelial carcinoma are also present in this cluster. A merged display of two clusters (e.g., clusters 6 and 7 or clusters 7 and 8) may show functional-related tissue areas in the broader sense. Cluster 9 is not informative and does not have any obvious histopathological interpretation. Cluster 10 (dark blue in Fig. 2b) displays preexistent connective tissue, small parts of normal squamous epithelium and mild epithelial dysplasia.

Fig. 2.

a Microscopic image of a larynx carcinoma section stained with H&E after MALDI acquisition. Regions of interest are manually annotated. b Segmentation map of 10 clusters generated from MALDI-imaging data of the section shown in panel A colors are arbitrarily assigned and are not interrelated with colors used in panel A (color figure online)

Fig. 3.

Automatically generated segmentation map and associated histological features of a larynx carcinoma section. Clusters 3–8 of overall 10 are displayed. a Segmentation map with the corresponding cluster visualized in black. b, c Superimposition of segmentation map and microscopic image of the H&E-stained section. The segmentation map was set half-transparent in order to evaluate its correspondence to the histological annotation. The borders between different clusters of the segmentation were linearly interpolated for better visualization. d Histopathological description of matching morphological feature (color as in segmentation map, Fig. 2) (color figure online)

Figure 4 shows m/z-images co-localized with cluster 3, for example, signals with high intensity inside the cluster spatial mask and low intensity outside. Cluster 3 represents dissociated growing tumor cell complexes characterized by clear cells and homogenous eosinophilic cell groups. Figure 5 shows the mean spectrum for cluster 3 (averaged over all pixels of this cluster) and indicates peaks of co-localized m/z-values plotted as images in Fig. 4. In accordance with our previous work, we assume that the signal at m/z 10,806.3 represents S100 A8 (also known as calprotectin), a protein previously shown to be characteristic for the transition from severe dysplastic tissue to cancer (Driemel et al. 2010). The molecules at m/z 3365.4 and 3479.4 are likely to represent alpha-defensins such as HNP1/DEFA1 (UniProt P59665) and HNP3/DEFA2 (UniProt P59666).

Fig. 4.

The spatial mask corresponding to the third cluster (first subplot) and 23 co-localized m/z-values. The cluster represents dissociated growing tumor cell complexes, characterized by clear cells and homogenous eosinophilic cell groups; see Fig. 3. Images of m/z-values are sorted by correlation with the spatial mask (first subplot), for each image its m/z-value is shown. Each m/z-image is visualized with the gradient color map with blue color corresponding to low intensity and red color corresponding to high intensity (color figure online)

Fig. 5.

The mean spectrum for the third cluster. The peaks of 23 co-localized m/z-values plotted as images in Fig. 4 are marked with red triangles (color figure online)

Discussion

The development of temporally- and spatially resolved proteomics techniques to study biological processes relevant to human health is one of the most important fields of current proteomic research. Temporal proteomics studies, that is, analyzing different time points of developmental processes, are difficult, especially in complex systems such as higher mammals, but analysis and visualization of protein distribution in complex tissues is possible. Traditional techniques depicting tissue morphology and localizing proteins in tissue are histological staining (e.g., H&E) and IHC. These techniques offer relatively high spatial resolution limited primarily by of the resolution of optical microscopy and can provide sub-cellular information, revealing protein localization in subcellular compartments (i.e., nuclei, cytoplasm or extracellular matrix). IHC is well established in routine pathology, allowing specific detection of known proteins if antibody probes are available. Consequently, the elucidation of new diagnostic or therapeutic markers by IHC is limited to applying existing antibodies to new sample sets. For a more global, unbiased detection of new markers, other techniques have to be employed. Such discovery techniques are DNA-microarrays in genomics, and various differential proteomic approaches combining different separation methods (e.g., 2D Gels, chromatographic separation) and protein identification by MS. However, these techniques are hampered by the fact that information about the spatial localization of proteins is oversimplified or not considered at all. This can produce misleading results, obscuring potential markers, especially those with specific but highly localized expression limited to relatively small tissue areas. Such markers may be highly useful, but will likely be obscured in analyses, which de-localize and proteins thereby average protein concentration (e.g., tissue homogenization for protein extraction).

To some extent, laser-capture microdissection of tissues alleviates these limitations by excising defined tissue features or cell types for further analytical steps. This methodology can be combined with SELDI-MS, which utilizes chips with affinity surface coatings. This technique shows a higher sensitivity of a broader mass range than comparable MALDI systems (Semmes et al. 2005). Several studies showed the potential of the combination of laser-capture microdissection with Protein Chip technology to analyze detailed regions of tissue sections and to identify several markers in diverse tumor entities (von Eggeling et al. 2000, 2001; Cazares et al. 2002; Jr et al. 1999). The combination of Protein Chip Arrays and laser-based microdissection, however, does not allow the analysis of the functional heterogeneity in a tissue in detail. Moreover, selection of the excised tissue is always guided by visible features (e.g., cell morphology) and is therefore never unbiased. Visually homogenous tissue with underlying molecular differences cannot be resolved.

With imaging mass spectrometry, it is possible to generate molecular profiles of proteins or other analytes directly from the tissue surface of thin sections (Seeley and Caprioli 2008). This technique analyzes morphologically intact tissue sections, avoiding homogenization and separation steps. Each of the molecular species detected in a MALDI-imaging data set can be used to generate one image in a form of molecular histology. These molecular images can be merged to analyze protein co-localization or superimposed with histological images (e.g., H&E or IHC stained tissue) to demonstrate the heterogeneity of such tissue samples. Complex correlations of protein expression can be visualized in these experiments and may indicate previously unknown co-localization of proteins, which may indicate their functional relationships. Because of the hundreds of molecular images generated by a single MALDI-imaging study and the multitude of possible correlations, the interpretation of such experiments is complex. Furthermore, it requires a strong cooperation between analytical scientists, bioinformaticians and clinical researchers to generate and evaluate the results in the context of the underlying histology and their clinical background (Deininger et al. 2008). Previously (Wehder et al. 2010), we compared IHC, microdissection followed by SELDI and MALDI-imaging to highlight advantages and disadvantages as well as synergistic effects. We used consecutive tissue sections from head to neck cancer and HNP1-3 and S100A8 as exemplary proteins. With all three techniques, the proteins HNP1-3 and S100A8 could be detected and localized to nearly identical regions. For the present study, we have established such a cooperation and demonstrated how the state-of-the-art computational method, namely the recently proposed segmentation pipeline (Alexandrov et al. 2010) in combination with the improved peak picking (Alexandrov and Kobarg 2011) can reveal molecular structure of in a complex tissue specimen of a larynx carcinoma (Alexandrov et al. 2010; Alexandrov and Kobarg 2011).

Spatial segmentation of MALDI-imaging data

The computed segmentation maps correspond to the morphological composition of analyzed tissue. Moreover, a segmentation map highlights molecular similarity of morphological structures, which can lead to improved understanding of functional processes in tissue.

Comparing to standard histological tools such as H&E or immunohistochemical stainings, MALDI-imaging segmentation (1) takes into account a large number of protein species at a time (2) is not a targeted but a data-driven approach that finds regions of similar molecular composition, (3) represents the tissue with several colors. Thus, segmentation maps represent a functional proteomic topography on a molecular level, which cannot be reached by other methods.

Naturally, the interpretation of a segmentation map showing complex proteomic patterns in a single image and allocation of the derived segments to morphological features requires histological expertise and may be limited by the lateral resolution of the MALDI-imaging analysis.

State-of-the-art spatial resolution of MALDI-imaging is much lower than that of light microscopy. Resolutions of 20 μm have been reported for the analysis of proteins (Lagarrigue et al. 2011; Deutskens et al. 2011) and resolution as low as 3 μm has been reported for the analysis of lipids and small peptides using custom-built instruments (Rompp et al. 2011). Further improvement of the resolution in MALDI-imaging experiments is likely and will probably allow routine acquisition at low μm resolution in the near future. Improvements beyond this level, however, will pose a considerable challenge for instrument sensitivity. As pixel size decreases, the number of ions generated “per pixel” and thus the resulting signal intensity, limiting high-resolution experiments to highly abundant analytes. Aside from improvements in actual instrument sensitivity, this limitation could be overcome by increasing the ionization efficiency of the MALDI process, which is generally low. MALDI ionization is an incompletely understood process, especially in complex mixtures (such as a tissue proteome). Analytes compete for ionization, which affects individual signal intensities in nonpredictable ways, resulting in strong constraints for quantitative approaches. Also, ionization is dependent on the chemical nature of each analyte, with a general bias toward lower mass molecules. As a result, proteins of high molecular weight are typically underrepresented in MALDI-imaging experiments. Analysis of other high molecular weight compounds, such as DNA or RNA is even more challenging because of their lower ionization efficiency. In this regard, improved sample pretreatment and preparation procedures proved beneficial by selectively enriching for a specific class of analytes or decreasing background (van Remoortere et al. 2010). Matrix deposition itself represents a combination of analyte extraction from the tissue, lateral diffusion of analytes and matrix crystal formation, and as such is a limiting factor for both sensitivity and resolution. Laser focus size and precise sample positioning also pose a limit on resolution but these problems can be addressed by technological advances (Rompp et al. 2011; Holle et al. 2006) as well as computational approaches (Alexandrov et al. 2011; Jurchen et al. 2005).

Histopathological assessment of the segmentation map

In the case of the larynx carcinoma analyzed here, which contains different functional areas, we could demonstrate that the clusters highlight functional proteomic domains in accordance with the histopathology of the tissue. Moreover, the clusters show defined functional areas, which cannot be easily detected in the H&E-stained section (e.g., cluster 3). The clearly visible segments facilitate the detection of similar tissue areas enabling an exact and rapid orientation also in highly complex tissues (Fig. 2a). Clear cluster borderlines facilitate the orientation in consecutive nonstained tissue sections for tissue microdissection and further proteomic analysis.

Beneath these distinct results, some areas and borders might contain artifacts or point to previously unknown biological phenomena. Cluster 2 shows a small overlap with parts of differentiated cancer components on the left and right border of the cluster. Cluster 7 contains an invasion front indicating the participation of stroma in the malignant transformation of the epithelium and invasion of the cancer. Areas with epithelial, dysplastic and stromal cells, tumor cells growing into stroma, and the stroma itself may be merged in a common cluster (clusters 6, 7). This might be a hint to interaction of tissue components or the involvement of the stroma in the process of malignant transformation.

Molecular identification of m/z-values

After proteomic images are generated, it is crucial to identify peaks of interest, especially if mechanistic insight into disease development is desired above mere diagnostic classification (Melle et al. 2004). In principle, MALDI-imaging offers the possibility to directly identify peaks of interest by on-tissue tandem-MS analysis. In practical application, this is limited by low sample amounts and, even more so, by an extremely complex background of interfering ions. A more promising possibility is to use a top-down proteomics approach, in which the HPLC fraction containing the protein of interest is analyzed by electrospray MS.

Our segmentation approach elucidated three highly intense m/z-values signals distinctively associated with dissociated growing tumor cell complexes. The corresponding masses could be assigned to proteins (HNP1, HNP2, S100A8) identified by us in earlier studies. This assignment is possible by the specific spatial pattern of these m/z-values found using segmentation and correlation. It has to be mentioned that HNP1 +2 are also classifiers for other clusters (6 + 7). HNP1–3 are part of the alpha-defensin family and are normally synthesized in neutrophil precursor cells. Mature circulating neutrophils then release them at inflammatory sites (Albrethsen et al. 2005, Melle et al. 2005). In cancers like renal cell carcinoma (Muller et al. 2002), colorectal cancer, or oral squamous cell carcinoma (Lundy et al. 2004), HNP1–3 were also found to be upregulated. S100A8, also known as migration inhibitory factor-related protein (MRP)-8 or calgranulin A can form a heterodimer complex with S100A9 (MRP-14, calgranulin B) (Vogl et al. 2007). Both, but especially S100A8, are involved in the regulation of cell proliferation (Hermani et al. 2006) and acute inflammation and were found to be up- and downregulated in many cancers including gastric cancer (Yong and Moon 2007), prostate cancer (Hermani et al. 2005), head and neck cancer (Melle et al. 2004), breast carcinoma (Moog-Lutz et al. 1995), and colorectal cancer (Stulik et al. 1999). S100A8 is also involved in the metastatic process, acts as chemo-attractant for the homing of tumor cells to premetastatic sites and increases the motility of circulating cancer cells. In two of our previous studies with microdissected cancer tissue (Melle et al. 2004) and brush biopsies from oral cancer (Driemel et al. 2007, 2010) analyzed by SELDI, we showed an explicit and significant stepwise loss of S100A8 expression during the progression from normal to proliferative/inflammatory and finally to cancer cells.

In conclusion, this investigation shows that the segmentation provides a unique way to depict the complex functional proteomic heterogeneity of a tissue in one image. As a result, integral aspects of tissue function could be explored under diverse conditions like tumor proliferation, invasion, and pharmaceutical metabolisation. In addition, the clusters highlight functional borders for tissue microdissection, which is a useful tool for the subsequent identification of relevant markers. Further studies on cancer tissue samples cohorts will show whether segmentation maps are also helpful in the detection of new biomarkers for preliminary cancer and early invasion.