Cover Page. The handle holds various files of this Leiden University dissertation

Transcription

1 Cover Page The handle holds various files of this Leiden University dissertation Author: Lou, Sha Title: Biomarker discovery in high grade sarcomas by mass spectrometry imaging Issue Date:

2 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Sha Lou

3 Copyright 2017 Sha Lou. All rights reserved. No part of this thesis may be reproduced in any form or by any means without permission from the author. A catalogue record is available from the Leiden University Library ISBN: Printed by: Proefschriftmaken, Vianen, the Netherlands This research has been financially supported by COMMIT, Cyttron II and the ZonMW Zenith project.

4 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Proefschrift ter verkrijging van de graad van Doctor aan de Universiteit Leiden, op gezag van Rector Magnificus prof. mr. C.J.J.M. Stolker, volgens besluit van het College voor Promoties te verdedigen op dinsdag 16 mei 2017 om uur door Sha Lou geboren te Ningbo, China in 1984

5 Promotor: Copromotoren: Prof.dr. J.V.M.G. Bovée Dr.L.A. McDonnell Dr.B.D. Balluff (Maastricht University) Leden promotiecommissie: Dr. A.M. Cleton-Jansen Prof.dr. M. Wuhrer Prof.dr. G.L. Corthals (University of Amsterdam) Prof.dr. M. Clench (Sheffield Hallam University)

6 Contents Chapter Introduction Chapter 2 21 An experimental guideline for the analysis of histologically heterogeneous tumors by MALDI-TOF mass spectrometry imaging Biochim Biophys Acta Oct 8. pii: S (16) Chapter High-grade sarcoma diagnosis and prognosis: Biomarker discovery by mass spectrometry imaging Proteomics Jun;16(11-12): Chapter High nuclear expression of proteasome activator complex subunit 1 predicts poor survival in soft tissue leiomyosarcomas Clin Sarcoma Res Oct 1;6:17. Chapter 5 91 Prognostic metabolite biomarkers for soft tissue sarcomas discovered by mass spectrometry imaging J Am Soc Mass Spectrom Feb; 28(2): Chapter Summary, conclusion, discussion and future perspectives References. 123 Nederlandse samenvatting Curriculum vitae List of publications Acknowledgements.136

7

8 Chapter 1 Introduction

9 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Abbreviations CID: Collision-induced dissociation DESI: Desorption electrospray ionization ETD: Electron-transfer dissociation FFPE: Formalin fixed paraffin embedded FNCLCC: La Fédération Nationale des Centres de Lutte Contre le Cancer FTICR MS: Fourier transform ion cyclotron resonance mass spectrometry ITO: Indium tin oxide LC-MS/MS: Liquid chromatography tandem-mass spectrometry LMS: Leiomyosarcoma MALDI: Matrix-assisted laser desorption/ionization MFS: Myxofibrosarcoma OS: Osteosarcoma PSME1: Proteasome activator complex subunit 1 RF: Radio-frequency TMA s: Tissue microarrays TOF: Time-of-flight UPS: Undifferentiated pleomorphic sarcoma WHO: World Health Organization 2

10 Chapter 1 Introduction 1. Introduction 1.1 Clinically challenging high grade sarcomas A sarcoma (from the Greek word σάρξ sarx, meaning flesh ) is a rare kind of cancer that originates from mesenchymal cells, which are cells that form tissues such as bone and muscle. They can arise from all connective tissues in the body, including superficial as well as deep soft tissues, at the extremities or at the trunk, or in visceral organs. According to the 2013 World Health Organization (WHO) classification, sarcomas are histologically classified based on the normal tissue that they most resemble, for example sarcomas displaying smooth muscle differentiation are classified as leiomyosarcomas, while those that display osteogenic differentiation are osteosarcomas. High grade sarcomas lacking any line of differentiation are classified as undifferentiated pleomorphic sarcoma. (Fig.1.1) 1. Soft tissue sarcomas are more common than bone sarcomas and affect approximately 12,000 people each year in the United States 2. Besides being classified based on their line of differentiation, sarcomas are also divided in three histological grades predicting clinical behaviour. According to the most widely used grading system of La Fédération Nationale des Centres de Lutte Contre le Cancer (FNCLCC), sarcomas are labelled as low, intermediate or high grade based on three histological parameters including tumor differentiation, mitotic count, and tumor necrosis 1. Thus, histologically, high-grade sarcomas are less differentiated, have more mitoses and tumor necrosis and behave more aggressively. Fig.1.1 The main subtypes of high-grade sarcomas that are characterized by complex genotypes include leiomyosarcoma (LMS), myxofibrosarcoma (MFS), undifferentiated pleomorphic sarcoma (UPS) and osteosarcoma (OS). High-grade sarcomas constitute less than 1% of all cancers 1, but there are more than fifty histological subtypes with sometimes overlapping histological features 3. High-grade sarcomas can lack clearly defining features that indicate a line of differentiation at light microscopy (e.g. undifferentiated pleomorphic sarcoma). 3

11 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Additional complications may arise from the substantial heterogeneity often observed, in which tumor tissue may contain areas with different levels of differentiation, differing grades and cellularity, necrotic tissue, or inflammatory infiltrate, amongst others. Furthermore histologically identical regions may have different molecular content and differ in clinical behavior, e.g. response to therapy 4. The combination of rarity, many subtypes with overlapping and heterogeneous histologies, and molecular intratumor heterogeneity has made the correct diagnosis and treatment of high grade sarcomas challenging. This is reflected in the low concordance between primary diagnosis and second opinion, which has been determined to be just 56% 5. Consequently, there is an urgent need for new molecular signatures that could aid in a better discrimination between these poorly differentiated tumors and which could aid in predicting patient outcome. From the molecular point of view, there are two main types of sarcoma: sarcomas with simple genomes that are caused by a specific translocation or a specific gene mutation, and sarcomas that are characterized by complex karyotypes reflecting genetic instability. The latter category includes leiomyosarcoma, myxofibrosarcoma, undifferentiated pleomorphic sarcoma and osteosarcoma (Fig.1.1). They are all highly heterogeneous, may display pleomorphic histological features, and often have poor outcome 6. In leiomyosarcoma for example the 5- year survival rates range from 15%-60% depending on the location, size and stage of disease at time of diagnosis 1, and metastatic relapse is about 60% within five years after resection of the primary tumor 7. Surgery is the mainstay of treatment, and sometimes combined with radiotherapy or chemotherapy 8. Among leiomyosarcoma, myxofibrosarcoma, undifferentiated pleomorphic sarcoma and osteosarcoma, previous studies using gene expression and gene methylation have shown distinction as well as similarities. Discriminating genes between leiomyosarcoma and undifferentiated pleomorphic sarcoma have been reported, some of which were validated by quantitative PCR and immunohistochemistry. However, hierarchical clustering didn t show a clear separation between leiomyosarcoma and undifferentiated pleomorphic sarcoma 9. Carneiro et al. also reported similarities of both DNA copy number and gene expression signatures between leiomyosarcoma and undifferentiated pleomorphic sarcoma, which were not distinguishable by unsupervised hierarchical clustering 10. Based on gene methylation status, unsupervised clustering was able to distinctly separate sarcomas with specific translocations but could not distinguish between 4

12 Chapter 1 Introduction leiomyosarcoma, myxofibrosarcoma and undifferentiated pleomorphic sarcoma, which are characterized by complex genomes 11. The association of specific hypermethylated genes with neoplastic characteristics, metastasis and gender were also reported 12. The histological heterogeneity of high grade, complex karyotype sarcomas comprises a significant source of variability that may be countered by utilizing an imaging based analysis that can be integrated with histopathological analysis, such that specific histopathological regions may be compared. 1.2 Mass spectrometry imaging as a discovery tool Mass spectrometry imaging (also known as imaging mass spectrometry) is an analytical technique based on the spatially resolved mass spectrometric analysis of a surface, e.g. a biological tissue section. In a mass spectrometry imaging experiment a focused ionization beam is used to analyze the molecular content of a small, localized region of tissue. By repeating this process for an array of positions spread across the tissue the molecular content and distribution of the tissue may be investigated. To separate and detect molecules by mass spectrometry, the compounds first need to be liberated into the gas phase and ionized. Following that, they are separated on the basis of their mass/charge ratio (m/z) in the mass analyzer, before reaching the detector connected to the data storage system (Fig.1.2). Ion Source Mass Analyser Detector Data system vacuum vacuum Fig.1.2 Schematic components of a mass spectrometer Mass spectra Data analysis Different ionization methods have been developed and adapted for mass spectrometry imaging including secondary-ion mass spectrometry 13, laser desorption/ionization 14, matrix-assisted laser desorption/ionization (MALDI) 15 and several variations on desorption electrospray ionization (DESI) 16,17. There are also some emerging ionization methods showing promise including imaging mass cytometry 18 and laser ablation electrospray ionization 19. There have also been a 5

13 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging number of mass analyzer developments for mass spectrometry imaging, specifically to increase mass resolving power, mass accuracy, MS/MS capabilities and spatial resolution. The most common ionization techniques used today for MSI are MALDI and DESI, and the most common mass analyzers are time-of-flight (TOF) and Fourier transform ion cyclotron resonance mass spectrometry (FTICR MS). TOF: A time-of-flight (TOF) mass analyzer measures the time taken by ions to move from the ion source, through a fixed-distance flight tube and at a defined kinetic energy, to a detector (commonly a microchannel plate type detector). Ideally the time of flight is given by Formula A: t= k (m/z) k=d/ 2U where t is the time of flight; m/z is the mass to charge ratio; d is the length of the flight tube and U is the strength of the electric field through which the ions are accelerated. Formula A refers to a linear TOF analyzer; linear TOF s are high sensitivity instruments able to detect ions over a very wide mass range but with low mass resolution. It is for this reason that linear TOF s are extensively used for protein MSI, in which the molecular ions may span a very wide m/z range and average masses may be sufficient to assign identities. The mass resolution of the mass spectrometer, the sharpness of the peak detected by the mass spectrometer (Δm/m), defines the specificity of the peaks; i.e. lower resolution peaks may be assignable to more potential molecules. The mass resolution of a TOF analyzer may be increased by using a reflectron device, which is essentially an ion-mirror that is able to reflect ions onto a different path 20. A reflectron is often used to double the effective flight path, in which the ions follow a V-shaped path. The longer flight path increases mass resolution (Formula A), and significant gains are also provided by the time focusing afforded by the ion mirror (more energetic ions progress deeper into the mirror before being reflected, thus enabling minor differences in kinetic energy to be compensated) 20. However reflectron devices are typically only used for lower mass molecules. In summary, MSI linear TOF s are used for protein MSI, whereas reflectron TOF is used for the analysis of metabolites, drugs, lipids and glycans. FTICR: Fourier transform ion cyclotron resonance (FTICR) mass spectrometry is based on the circular, cyclotron motion of an ion in a magnetic field. Through the use of superconductig magnets, typically 7-15T, ultra high mass resolution and high mass accuracy may be achieved. This enables MSI analysis of peaks that are 6

14 Chapter 1 Introduction highly specific to single species, which may be assigned on the basis of high mass accuracy 21. At room temperature an ion in a magnetic field moves in a very small orbit at its cyclotron frequency, e.g. an ion of m/z 100 in a 12T magnetic field will have an ion cyclotron radius of 20 m. Such a small orbit cannot be used for ion detection. However applying a radio-frequency (RF) electric field at the ion s cyclotron frquency increases the ion cyclotron radius 21. After excitation, all ions of the same m/z will move as a coherent ion packet, the movement of which can be detected via the image charge they create on detector plates. Sweeping the RF field through a range of cyclotron frequencies excites ions of different m/z; the resulting signal consists of the sinusoidal response of each m/z present in the ICR detector cell, superimposed onto the detection plates 21. Application of the Fourier transform to this time domain signal results in a frequency domain mass spectrum, in which peaks are detected at their ion s cyclotron frequency (which is dependent on the m/z of the ion). An experimental calibration using ions of known m/z is then used to convert this frequency domain signal into an accurate-mass spectrum 21. Among all mass spectrometry imaging systems MALDI-TOF is the most widely used, owing to its high speed, high sensitivity and parallel detection over a wide mass range. These characteristics have also made it the most commercially available, and for which user-friendly data acquisition and data analysis software has been developed. It is currently the approach most widely used for clinical research 22. A typical workflow of a MALDI-TOF mass spectrometry imaging experiment is as follows: a μm thick tissue section (fresh frozen tissue or formalin fixed paraffin embedded tissue) is cut using a (cryo)microtome and placed onto an indium tin oxide coated glass slide (the ITO slide is conductive and transparent, enabling MALDI-MSI and histology of the same tissue section). The tissue is then washed to remove cofactors present in the tissue that would otherwise adversely affect measurement sensitivity, e.g. salts. Then the matrix, typically a small organic molecule, is applied onto the tissue section by either spraying, the deposition of discrete droplets, or sublimation. The matrix solution dissolves the biomolecules present in the tissue; upon further evaporation of the matrix solvent the tissue s biomolecules are extracted to the surface where they are either incorporated into or remain in close contact with the crytalizing matrix crystals. After allowing the matrix coated tissue to dry, the matrix coated tissue section can be analyzed by MALDI mass spectrometry imaging. It should be noted that tissue washing and matrix deposition is critical to the success of a MALDI-MSI experiment. 7

15 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging When the laser hits the matrix coated tissue section, the matrix rapidly and efficiently absorbs the laser irradiation leading to the ablation and ionization of the incorporated biomolecules. The m/z and amount of gas phase biomolecular ions are then determined by the mass analyzer (Fig.1.3a). In MALDI mass spectrometry imaging the biomolecular ions are typically singly charged, through the gain or loss of a proton, which means that the molecule s mass can be easily calculated. In a mass spectrometry imaging experiment, a mass spectrum is recorded for each of the many thousands of pixels, each of which is considered to be representative of the complex biomolecular state of that localized region of tissue. The distribution of a detected molecular ion can be displayed by plotting its variation in intensity as a function of position. One of the distinct advantages of MALDI mass spectrometry imaging is that it does not harm the tissue s histology: after the MALDI-MSI experiment any remaining matrix may be removed from the tissue using an alcohol-based wash, and then the tissue histologically stained, e.g. with hematoxylin and eosin. The histological image and MALDI mass spectrometry imaging data can then be co-registered through the use of fiducial markers or using unsupervised registration techniques 23. Once registered a histopathological analysis may be performed and specific regions annotated on the basis of their histological features. These annotations enable the MALDI MSI spectra from histopathologically defined regions of interest to be exracted (a so-called virtual microdissection). The determination of biomolecular ions specifically associated with distinct histopathological entities is widely used to identify biomarkers for diagnosis or differential diagnosis 24. In cancer research, tumor specific mass spectra can be tested for correlation with the clinical outcome from the patient data to identify mass spectrometric signatures associated with prognosis or response to therapy (Fig.1.3b). Compared to other techniques for large-scale protein separation and identification, mass spectrometry imaging doesn t require tedious and time-consuming extraction and fractionation steps; it also retains the original spatial relationship among the different molecules. Established techniques like immunohistochemistry can also achieve the same goal, but mass spectrometry imaging has higher throughput (multiple molecules can be analyzed simultaneously) and doesn t need prior knowledge in the form of labeling of a certain compound of interest. Because of the seamless integration of spatially resolved mass spectrometric data 8

16 Chapter 1 Introduction with histopathological knowledge, mass spectrometry imaging has been rapidly adopted for clinical research for biomarker discovery 22,25,27,28 and the direct molecular assessment of tissue samples (Fig.1.4), especially for heterogeneous tumor tissues where molecular features might not be mirrored by histology. Below it is described how MALDI mass spectrometry imaging has been applied in clinical research. Fig.1.3 (a) A typical workflow of a MALDI mass spectrometry imaging experiment. A tissue section is prepared by cryo-sectioning and matrix deposition; a spatially resolved MALDI MS analysis of the tissue generates a data cube of pixel correlated mass spectra (reprinted with permission from McDonnell et al. 25. Copyright 2010 Elsevier B.V.). (b) Histology guided MALDI mass spectrometry imaging data analysis: following removal of any remaining matrix the tissue section is histologically stained, then histopathologically 9

17 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging examined to define regions-of-interest; the average mass spectrum from each histopathologically defined regions of interest can then be compared to identify diagnostic biomarkers, or compared with clinical outcome to identify prognostic biomarkers. Fig.1.4 Mass spectrometry imaging offers enormous potential to complement current histopathological methods as it enables the spatial visualization of biomolecules within their morphological context (reprinted with permission from Rauser et al. 26 Copyright 2011 Springer-Verlag, Ltd.) Diagnostic biomarker discovery The seamless integration of mass spectrometry imaging and histolopathology enables cell-type-specific mass spectral profiles to be extracted from patient tumor samples; by comparing changes in signal intensities associated with a disease for a large number of patient samples, the statistical significance of the molecular ions may be assessed. Candidate biomarkers that exhibit statistically significant changes in abundance should then be validated using independent tissue samples or independent methods. In the first report of the clinical application of MALDI MSI Stoeckli et al. 29 alerted researchers to its ability of distinguish cell-type-specific proteins: the protein thymosin β4 was found to be primarily located in the proliferating head of the tumor whereas the protein S100A4 was found localized in the tumor core. Subsequent studies have futher demonstrated this ability by revealing tumor type specific proteins 30,31, including with independent validation by immunohistochemistry 32. In breast cancer, specific peptides and proteins were found that were asscociated with HER2 status and which could be used to classify 10

18 Chapter 1 Introduction breast cancer patients 33. It was then demonstrated that this same breast-cancer derived proteome classifier could also predict HER2 status in gastric cancer 34. Formalin fixed paraffin embedded (FFPE) tissues represent, by far, the most common format for clinical tissue banks. The protein network formed by formalin prevents the analysis of intact proteins, but the combination of antigen retrieval and on-tissue proteolytic digestion was found to enable protein analysis (through their proteolytic peptides); and the tryptic peptide signatures could also act as biomarkers to differentiate different tissue types 35. Tissue microarrays (TMA s) contain FFPE needle core biopsies from many patient samples within a single block, and are invaluable resources for quickly assessesing the clinical significance of candidate biomarkers. Groseclose et al. reported the MALDI mass spectrometry imaging analysis of a TMA containing 112 lung cancer patients and showed protein signatures that could distinguish between adenocarcinoma biopsies and squamous cell carcinoma biopsies 36. MALDI analysis of lipids and of intact proteins has been shown to be able to differentiate lung cancer according histologic type 37 and to identify signatures for distinguishing primary tumors from metastases and patients with good prognosis from those with poor prognosis 38. In ovarian cancer MALDI mass spectrometry imaging has revealed protein signatures that discriminate between benign and malignant tumors 39,40. Longuespee et al. 41 then demonstrated that one of these protein biomarkers (Reg alpha) could be used for early diagnosis and tumor-relapse. The first clinical study of human prostate cancer by MALDI mass spectrometry imaging was from Schwamborn et al. 42 who detected and identified overexpressed proteins for non-cancerous glands and cancerous glands. Following which Cazares et al. 43 revealed, identified and validated a tumor specific biomarker based on the analysis of a large patient series of tissues and which represented the first large scale clinical investigation with independent verification. In other types of cancer, recent work from Pagni et al. identified specific protein signatures distinguishing between malignant and benign lesions on thyroid carcinomas (fresh tissue sections collected within 15min after surgery) followed by assigning additional samples which were confirmed by morphology 44. Besides lipids, peptides and proteins, it was demonstrated that MALDI mass spectrometry imaging can also be used to analyze metabolites and glycans, and that these molecular classes also contained biomarkers that distinguish between normal and tumor tissues. For example the recent analysis of a TMA containing 11

19 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging chromophobe renal cell carcinoma and renal oncocytoma found 123 metabolite peaks that exhibited significant differences in intensity 45 ; similarly specific glycans and glycan signatures have been reported that differentiate pathological tissue from healthy tissue for multiple cancer types 46,47 and osteoarthritis Prognostic biomarker discovery The mass spectrometry signatures obtained from MALDI mass spectrometry imaging may be compared with clinical endpoints such as overall survival, metastasis-free survival, or response to therapy. Early work from Schwartz et al. 49 and Yanagisawa et al. 50 provided the first evidence that MALDI mass spectrometry imaging could correlate mass spectrometric profiles to patient survival. These may be single proteins or protein signatures; an example of the latter was provided by Balluff et al. 51 who reported a prognostic seven-protein signature composed of novel tissue markers in intestinal-type gastric cancer. Aichler et al. 52 reported proteins associated with response to neoadjuvant chemotherapy in Barret s adenocarcinoma; it was found that the mitochondrial protein COX7A2, as well as several other mitochondrial proteins essential for the electron transport chain of oxidative phosphorylation, was significantly lower in patients that responded to chemotherapy. It was postulated and proven that the lower levels of mitochondrial proteins was due to mitochondrial abnormalities, and which imparted increased sensitivity to cisplatin treatment. This manuscript was the first example in which MALDI mass spectrometry imaging led to the discovery of biomarkers, and the real biological mechanism behind the different patient response/survival. The fact that COX7A2 was also found to be a prognostic marker in Barrett s adenocarcinoma 53, demonstrates how MALDI mass spectrometry imaging can aid patient management, by identifying those at higher risk and indicating those who will respond to therapy Small molecule MALDI mass spectrometry imaging MALDI mass spectrometry imaging is able to analyze different molecular classes by using different tissue preparations, specifically different tissue washes and different matrix solutions. Using the matrix 9-aminoacridine, lipids and metabolites such as adenosines may be analyzed 54,55 ; with on-tissue derivatization strategies this has been expanded to include amino metabolites and neurotransmitters 27,56. 12

20 Chapter 1 Introduction Metabolite MSI offers enormous clinical potential by enabling the imaging of a largely previously intractable class of biomolecules and, if combined with known metabolic pathways, provide a means to image the activities of the pathways in tissues. Altered metabolism is now a hallmark of cancer 57. Dekker et al. used 9- aminoacridine as matrix resulting in the direct detection of metabolites from oncocytic follicular thyroid cancer and breast cancer tissues 58. Many endogenous metabolites were simultaneously recorded including pyruvate, lactate, fumarate, succinate and metabolites associated with Warburg effect 59. Walch and workers have demonstrated that FFPE tissues retain many metabolites, which can act as both diagnostic and prognostic biomarkers Identification of biomarker ions The identification of the biomarkers found by MALDI mass spectrometry imaging experiment is important for several reasons: i. It may translate into biological insights, as exemplified by COX7A2 in Barret s adenocarcinoma example referred to above. ii. As it is an unbiased approach, it will help guard against false positives. Without knowledge of the identity of the biomarker species, one cannot independently verify them and so remains vulnerable to unknown sources of bias. In general, the direct identification of ions detected by MALDI mass spectrometry imaging is non trivial. A major factor is the absence of a purification step of the single molecules, but there are also factors that are specific for the respective molecular classes. Metabolites for instance lack the necessary MALDI fragment databases for identification. In contrast, intact proteins are too large to be efficiently fragmented by collision-induced dissociation (CID) and their single charge states are incompatible with electron-transfer dissociation (ETD). Instead the spatially resolved MALDI mass spectrometry imaging data is matched to the proteins identified by liquid chromatography tandem-mass spectrometry (LC- MS/MS) analysis of non-spatially resolved protein extracts. The proteins may be identified via their proteolytic peptides (bottom-up proteomics, the most common strategy) or direct LC-MS/MS analysis of the intact proteins (top-down proteomics, less common). Identification of proteins detected by MALDI mass spectrometry imaging A bottom-up protein identification experiment consist of five stages 60, 1) protein 13

21 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging extraction of tissues; 2) protein digestion by an enzyme, typically trypsin; 3) peptide separation and purification mainly using liquid chromatography, and ionization by electrospray; 4) MS spectra acquisition; 5) MS/MS spectra acquisition from peptides of interest including isolation, fragmentation, and mass analysis. Proteins are identified through a statistical comparison of the experimental m/z of the precursor ions and fragment ions, with those obtained via the in-silico digestion and fragmentation of proteins in a curated database 61. Top-down based protein identification is essentially the same process but without the proteolytic digestion, and in which the liquid chromatography and MS/MS must be adapted for the larger size of intact proteins (and are, in general, more challenging) 62,63. The proteins or proteolytic peptides detected in a MALDI mass spectrometry imaging experiment are assigned the identities of proteins identified by LC-MS/MS on the basis of mass accuracy (i.e. how well the protein measured by MALDI mass spectrometry images matches the mass of the identified protein). A significant number of clinical MALDI mass spectrometry imaging studies have been reported in which the proteins detected were assigned identities, and are now available as public databases where users can readily retrieve previous assignments of similarly sized proteins detected by MALDI mass spectrometry imaging. The mass matching at the heart of these assignments of protein identities are highly dependent on mass accuracy, and so additional constraints such as isotope pattern have been utilized 67. Nevertheless the approach has proven reasonably successful, much more than would be expected on the basis of the very large number of proteins and possible protein isoforms present in an organism. The reason it is so successful is the limited dynamic range of a MALDI mass spectrometry imaging experiment: while an indepth LC-MS/MS based analysis of a tissue section may quantify protein groups, they are based on a very large number of cells, typically greater than In contrast each pixel s mass spectrum in a MALDI mass spectrometry imaging experiment utilizes between 1 and 25 cells, depending on the spatial resolution of the experiment. Accordingly a MALDI mass spectrometry imaging experiment can only analyse the more abundant proteins; mass matching is effective because the number of accessible proteins is much smaller than the entire proteome 64. Note: it is a testament to the sensitivity of the analysis that MALDI mass spectrometry imaging can simultaneously image several hundred proteins. When reporting biomarkers it is important to independently validate the finding; this is especially true when the 14

22 Chapter 1 Introduction identity is assigned by mass matching. Identification of metabolites detected by MALDI mass spectrometry imaging A number of publicly accessible databases, such as the Human Metabolome Database ( and Metlin (metlin.scripps.edu), are available that can be searched on the basis of the peaks detected in a metabolite mass spectrum. These databases included extensive information, ranging from elemental composition, MS/MS spectra and membership of metabolic pathways. Accordingly the databases can be used to assign metabolite peaks on the basis of accurate mass matching, e.g. <1 ppm mass accuracy, and assignments further confirmed by comparing isotope patterns or MS/MS spectra (Supervised) tissue classification MALDI mass spectrometry imaging generates rich datasets containing hundreds molecular features. Supervised classification 68 uses prior knowledge about the tissue samples to either identify latent variables in the data that can distinguish different patient groups (e.g. differential diagnosis, differential prognosis), or to generate a model that can be used to classify samples of unknown origin. A classification algorithm is first trained and cross-validated using well-defined patient groups, e.g. tumor tissue versus healthy controls, and then validated on independent sets of well-defined patients 69. Rauser et al. 70 and Balluff et al. 71 have reported MALDI mass spectrometry imaging classification models to distinguish HER2 positive and HER2 negative tumors that could determine HER2 status in both breast (accuracy: 89%) and gastric (accuracy: 90%) cancer patients, respectively. This approach has also been used to differentiate primary tumors from secondary tumors (which maybe of unknown origin): a classifier trained using 17 peptide peaks detected from primary breast and pancreatic cancer samples was able to discern primary and metastasized breast and pancreatic samples 72. In MSI, supervised classification can be used for the classification of patients or of each pixel within the image (i.e. spatially resolved classification). The former uses the average mass spectrum from a (virtually microdissected) region to classify a patient, whereas the latter uses the per-pixel spectra for classification. The application of supervised classifiers to tumor samples on a pixel-level has revealed 15

23 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging a high degree of intratumoral molecular heterogeneity in prostate cancer 42, myxofibrosarcoma 4,73, and brain tumors Intratumoral heterogeneity investigation Clonal evolution, microenvironmental stress and multilineage differentiation of cancer stem cells can result in intratumoral heterogeneity 75,76. Intratumoral heterogeneity introduces substantial complexity into tumor research because different tumor cell populations may have distinct molecular phenotypes and differ in clinical behavior 77 (Fig.1.5a). By comparing the similarity of the spatially correlated mass spectra, mass spectrometry imaging has been used to further discriminate histologically homogeneous but molecularly different tumor areas. Deininger et al. 78 were the first to postulate that MALDI mass spectrometry imaging may be able to uncover tumor subpopulations within histologically uniform areas. They reported a hierarchical cluster analysis of intestinal type gastric cancer; the tumor was separated from the healthy tissue in one of the first branches of the dendrogram, subsequent branches of the dendrogram revealed a patchwork of molecularly distinct regions that were speculated to reflect the clonal composition of the tumor. The same tools were then used to reveal highgrade-like and low-grade-like regions in multiple patient tissues of intermediate grade myxofibrosarcoma, which could then be further subdivided using a classification approach 4. The crucial question when analyzing heterogeneity is whether the different subpopulations impact the patient. Balluff et al. have demonstrated that the molecularly distinct subpopulations revealed by MALDI mass spectrometry imaging may be statistically associated with patient outcome: specifically, discrete tumour subpopulations were revealed to be associated with overall survival in gastric cancer patients; and to the presence of locoregional metastases in breast cancer patients 79, Fig. 1.5b. 16

24 Chapter 1 Introduction Fig.1.5 (a) Tumor cell evolution results in different clones, some of which develop the capacity to metastasize, resist chemotherapeutic intervention or adversely affect the outcome of the patient. (b) Identification of clinically relevant tumor clones was performed by testing the statistical association between the presence of a specific subpopulation and patient outcome. First the contribution of each subpopulation, to each patient tissue sample, was determined. The clinical data from each patient was then assigned to the subpopulations, and then used to determine the phenotypic contribution of each subpopulation. Reprinted with permission from Balluff et al. 79. Copyright 2014 Pathological Society of Great Britain and Ireland. Published by John Wiley & Sons, Ltd.) In summary, mass spectrometry imaging is a powerful tool to study tumors, because the combination of spatially resolved mass spectrometry and histopathology of the same tissue section enables the molecular signatures of distinct cell types to be obtained from within the correct histopathological context of real patient tissues. Such specificity is essential for the analysis of molecularly and histologically heterogeneous tumors such as high-grade, complex-karyotype sarcomas Application of MALDI mass spectrometry imaging to sarcomas MALDI mass spectrometry imaging has been previously used to investigate sarcomas. The first investigation was reported by Caldwell et al. 80 who investigated the tumor borders of high-grade malignant fibrous histiocytoma (new WHO 17

25 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging classification has renamed this entity as undifferentiated pleomorphic sarcoma 1 ) and revealed a gradient of expression of tumor associated proteins, acyl- CoA binding protein (m/z 9,910), MIF (m/z 12,338) and histone H4 (m/z 11,385), into the surrounding histologically normal tissue. Willems et al. 4 reported tissue-type and tissue-grade specific proteins and lipids for differentiating myxofibrosarcoma and myxoid liposarcomas; the changes in lipid composition detected between lowgrade and high-grade myxoid liposarcoma were also consistent with the known biology of the tumor. The protein CRIP1 was first identified by MALDI mass spectrometry imaging as a prognostic biomarker for gastric cancer 51. CRIP1 immunohistochemistry then demonstrated that it was also correlated with favorable outcome and less metastases in osteosarcoma 81 and metastasis-free survival in breast cancer. 82 MALDI mass spectrometry imaging has also been used to investigate the substantial intratumoral heterogeneity present in soft tissue sarcomas; the application of semi-supervised hierarchical clustering 4 or other multivariate data analysis techniques 73 to MSI datasets revealed areas of tissue with distinct mass spectral signatures. This work together with that of others demonstrated the ability of mass spectrometry imaging to uncover molecular variation in histologically homogenous tumor areas. MALDI mass spectrometry imaging has also been used to investigate drug penetration into animal models of sarcoma, and the protein response of a sarcoma to therapeutic intervention. Huber et al. 83 investigated the distribution of the drug sorafenib in a sarcoma (A673) xenograft and demonstrated higher drug levels in regions with higher vascularization. Disrupting the tumor vasculature is an established cancer therapeutic strategy because selective cessation of tumor blood flow is known to lead to tumor cell death. Cole et al. 84 have used MALDI mass spectrometry imaging to investigate how the tumor in a mouse fibrosarcoma model responds to vascular disrupting agents, the aim being to identify potential resistance pathways. The specific strength of MSI is its seamless integration with histopathological analysis, enabling mass spectrometric signatures to be obtained from specific groups of cells even from within morphologically heterogenous tumor tissues. In this thesis the potential of MSI to identify biomarkers from high grade sarcomas, tumors characterized by their high heterogeneity, was investigated. 18

26 Chapter 1 Introduction 1.3. Objective, main results and content of this thesis Objective and study design High-grade sarcomas can be diagnostically challenging due to the combination of rarity, overlapping histology and substantial histological and molecular heterogeneity. For oncologists, sarcomas can be challenging due to a sometimes only moderate or lack of response to chemo/radio-therapy. But as histologically identical different subtypes may differ in clinical behavior and require different patient management, discriminating between subtypes becomes crucial. In this study, four subtypes of high-grade, complex-karyotype sarcomas were investigated. They are leiomyosarcoma, myxofibrosarcoma, undifferentiated pleomorphic sarcoma and osteosarcoma. They are all histologically heterogeneous, genetically complex, and often have poor patient outcome. I applied MALDI mass spectrometry imaging to address the following technical and scientific questions: 1) Can a general MALDI mass spectrometry imaging workflow be developed for high-grade sarcomas? What is the optimized protocol? 2) Are there MALDI-mass spectrometric molecular signatures that distinguish between high-grade leiomyosarcoma, high-grade myxofibrosarcoma, undifferentiated pleomorphic sarcoma and high-grade osteosarcoma that could serve as diagnostic biomarkers? 3) Are there MSI profiles indicative of patient prognosis? Can we identify prognostic biomarkers? 4) Does the degree of differentiation or tumor grade impact the ability to identify diagnostic and prognostic biomarkers? 5) Is undifferentiated pleomorphic sarcoma a common end-stage of complexkaryotype sarcomas that reflects the loss of differentiation i.e. high grade leiomyosarcomas with loss of myogenic differentiation; high grade myxofibrosarcomas with loss of myxoid areas; high grade osteosarcomas with loss of osteoid differentiation? Is MALDI mass spectrometry imaging able to identify molecular signatures within undifferentiated pleomorphic sarcoma that indicate its origin? Main content of this thesis The overall aim of the studies was to use optimized mass spectrometry imaging methods to discover and identify molecular biomarkers in high grade sarcomas. 19

27 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Novel biomarkers could improve the diagnosis and predict prognosis in these challenging tumours. In Chapter 2 multiple aspects of the MALDI mass spectrometry workflow are optimized for the analysis of high-grade sarcomas, ranging from the tissue preparation and data acquisition protocols, to the post-msi histological staining method, data quality control, histology-guided data selection, data processing and statistical analysis. Chapter 2 also describes the development of classifiers to relate the molecular signatures obtained from undifferentiated pleomorphic sarcoma to those obtained from high-grade leiomyosarcoma, high-grade myxofibrosarcoma, and high-grade osteosarcoma. In Chapter 3 we perform a biomarker discovery investigation, in which diagnostic and prognostic protein biomarkers are determined. Owing to the high degree of histological and molecular heterogeneity of high-grade sarcomas, the study also addresses how the biomarkers are affected by tumor grade/degree-of-cellulardifferentiation and if tumor subpopulations are present that are statistically associated with patient survival. Note: A selection of the patient tissue blocks that were used for biomarker discovery were first used to optimize the method. Note: the optimization was performed prior to the biomarker discovery experiments, even though the methodology paper was published afterwards. Furthermore, the MSI optimization only concerned the quality of the mass spectral information (i.e. number of mass spectral peaks, MS image quality, robustness of method), at no point was diagnostic or prognostic information used during the optimization process. Accordingly there was no possibility for the optimization to lead to systematic bias. In Chapter 4 we validate one of the prognostic protein ions, proteasome activator complex subunit 1 (PSME1), in an independent sample set, using immunohistochemistry on a tissue microarray containing a large patient series of high-grade soft tissue sarcomas. Chapter 5 describes a biomarker discovery investigation in which prognostic metabolite biomarkers are determined, and identified using high mass accuracy, ultra-high mass resolution MALDI mass spectrometry imaging on a 9.4T Fourier transform ion cyclotron resonance mass spectrometer. In Chapter 6 the results are summarized and future perspectives are discussed. 20

28 Chapter 2 An experimental guideline for the analysis of histologically heterogeneous tumors by MALDI- TOF mass spectrometry imaging Lou S, Balluff B, Cleven AH, Bovée JV, McDonnell LA. Biochim Biophys Acta Oct 8. pii: S (16) doi: /j.bbapap

29 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Abstract Mass spectrometry imaging (MSI) has been widely used for the direct molecular assessment of tissue samples and has demonstrated great potential to complement current histopathological methods in cancer research. It is now well established that tissue preparation is key to a successful MSI experiment; for histologically heterogeneous tumor tissues, other parts of the workflow are equally important to the experiment s success. To demonstrate these facets here we describe a matrix-assisted laser desorption/ionization MSI biomarker discovery investigation of high-grade, complex karyotype sarcomas, which often have histological overlap and moderate response to chemo-/radio-therapy. Multiple aspects of the workflow had to be optimized, ranging from the tissue preparation and data acquisition protocols, to the post-msi histological staining method, data quality control, histology-defined data selection, data processing and statistical analysis. Only as a result of developing every step of the biomarker discovery workflow was it possible to identify a panel of protein signatures that could distinguish between different subtypes of sarcomas or could predict patient survival outcome. Key words biomarker discovery; mass spectrometry imaging; high grade sarcoma; protocol optimization Abbreviations MSI: Mass spectrometry imaging; MALDI: Matrix-assisted laser desorption/ionization; OS: Osteosarcoma; LMS: Leiomyosarcoma; MFS: Myxofibrosarcoma; UPS: Undifferentiated pleomorphic sarcoma; S/N: Signal to noise ratio; H&E: Hematoxylin and eosin; ROI: Regions of interest; ANOVA: Analysis of variance; SVM: Support vector machine 22

30 Chapter 2 Methodology 2. An experimental guideline for the analysis of histologically heterogeneous tumors by MALDI-TOF mass spectrometry imaging 2.1 Introduction MALDI MSI and study objective Mass spectrometry imaging (MSI) has been rapidly adopted for clinical research because of its ability to seamlessly integrate spatially resolved mass spectrometric analysis with histopathological knowledge 28,85,86. Biological tissue analysis by MSI was first reported by Caprioli et al. in Since then MSI methods have been developed for the analysis of intact proteins 29, lipids 88, metabolites 89, neurotransmitters 27, glycans 90 and pharmaceuticals 91. One of the greatest opportunities and challenges for MSI lies in its clinical potential to aid patient diagnosis, prognosis, and therapy response through the molecular assessment of tissues obtained from patients 86. Many different ionization techniques have been developed or adapted for MSI including secondary-ion mass spectrometry 92, matrix-assisted laser desorption/ionization (MALDI) 93 94, desorption electrospray ionization and imaging mass cytometry 95. To date the availability of MALDI instruments from multiple vendors, as well as training courses focused on MALDI MSI, has made it the method most frequently used for clinical research 22. Much of the focus of clinical MALDI MSI is on cancer research as it enables, through an integration of the histological information, a differentiated analysis of the tumor tissues 24,96,97. This is especially pertinent for histologically heterogeneous tumor tissues, in which the normal tissue, tumor, stroma, and inflammatory infiltrate often present in patient tissues can each generate different molecular signatures. Such histological complexity/heterogeneity combined with the need in biomarker discovery experiments to compare histologically comparable tissues from multiple patients, by necessity places histopathology at the forefront of the data analysis pipeline. The clinical application of MALDI MSI is highly multidisciplinary and involves multiple distinct expertises; including tissue preparation, MS data acquisition, histopathology, data quality control, mass spectral data processing and statistical analysis. There is a recognized need in the MSI field to integrate these expertises in form of standardized MSI methods, data formats and reporting guidelines so that MSI-based assays may be 23

31 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging transferrable to other laboratories 98. Here we provide detailed experimental guidelines on how to use MALDI MSI to analyze clinically challenging tumors, in order to stimulate the field to develop standard operating procedures for clinical applications High grade sarcomas and study design Soft tissue sarcomas are a relatively rare and heterogeneous group generally found difficult by pathologists because of their rarity, and the fact that >50 entities are described with overlapping morphology 1,3. Distinction is crucial since they differ in clinical behavior and require different treatment 100. Soft tissue sarcomas can be subdivided in two main categories, sarcomas with simple genomes that are caused by a specific translocation or a specific gene mutation, and sarcomas that are characterized by complex karyotypes reflecting genetic instability. High-grade leiomyosarcoma (LMS), high-grade osteosarcoma (OS), and high-grade myxofibrosarcoma (MFS), have complex karyotypes and can have highly heterogeneous histologies. Although leiomyosarcoma or osteosarcoma normally display myogenic or osteogenic differentiation, respectively, these areas are not always obvious and may even be absent in small biopsy specimens. Similarly, myxoid areas may be sparse to absent in high-grade myxofibrosarcoma. This loss of characteristic cellular differentiation has led to the hypothesis that undifferentiated pleomorphic sarcoma (UPS), a diagnosis of exclusion for tumors lacking a line of differentiation at the histological or immunohistochemical level, represents a common end-stage in which the undifferentiated part makes up the bulk of the tumor. Though the morphology may be very similar in certain tumor areas, previous studies provide ample demonstration of the ability of MSI to differentiate between microscopically identical tumors as well as reveal intratumor heterogeneity 78,79,101. Here we describe how MALDI MSI results can be used to identify biomarkers and to assess whether protein signatures were shared between UPS and high-grade LMS, MFS, or OS. This group of high grade sarcomas with complex karyotypes are those most frequently encountered in the clinic and are clinically challenging because of their moderate response to chemo- and radiotherapy. 24

32 Chapter 2 Methodology 2.2 Methods Sample collection Fresh frozen tumor samples from patients diagnosed with high grade LMS (n=12), MFS (n=13), UPS (n=13), and OS (n=16) were collected and handled as described previously 102. Each sample is from a unique patient. Available patient data included gender, age, neoadjuvant/adjuvant therapy status, length of follow up time and status, as was reported by Lou et al For convenience it is reproduced here in Supplementary Tab.2.1. All tumor samples were acquired during routine patient care. Note: patient inclusion criteria were defined separately for the differential diagnosis biomarker discovery and patient prognosis biomarker discovery experiments (Tab.2.2.), and so patient numbers differed for these analyses and from the total sample number. Patient selection criteria MSI quality control Differential diagnosis Patient prognosis Viable area 60% Sample source Primary/ recurrent/ metastatic Primary tumors Consistent diagnosis Yes No 1 Excluded spectra (%) <40% 2 Measurement bias Randomized measurement sequence Table 2.2. Sample inclusion criteria. 1 Note: One of the patient tumors was originally diagnosed as LMS (strong desmin staining) but upon recurrence histology resembled OS. This patient sample was excluded from the diagnostic analysis. 2 Note: 90% measured samples have less than 20% excluded spectra Sample revision Before sectioning for MALDI MSI all patient tissues were histologically reevaluated to ensure the tissue samples were representative of each sarcoma. From each patient s tissue block 5 µm thick tissue sections were cut and stained with hematoxylin and eosin (H&E), and then evaluated by an expert pathologist (JVMGB) according to the criteria of the 2013 World Health Organization criteria. MFS and LMS cases were histologically graded according to the La Fédération Nationale des Centres de Lutte Contre le Cancer. It was also required that all tissue samples contained more than 60% viable tumor, since widespread necrosis was expected in several OS samples due to neoadjuvant chemotherapy. 25

33 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging MALDI-MSI-histology data acquisition The protocol for a MALDI MSI analysis of proteins from fresh-frozen tissues is now well-established 103 and consists of: (i) cryo-sectioning thin tissue sections, ii) mounting of the tissue section on to an indium-tin-oxide coated slide, as this enables histological analysis before 104 or after MSI analysis of the tissue section 103 ; (iii) tissue washing, to remove compounds that adversely affect sensitivity, e.g. salts, small molecules, and lipids 105,106 ; (iv) matrix application, (v) MSI data acquisition, (vi) removal of excess MALDI matrix, (vii) hematoxylin and eosin (H&E) staining, (viii) digital scanning of the stained tissue, histological annotation by an expert pathologist, and co-registration to the MALDI MSI dataset 97, and (ix) data analysis based on selected regions of interest, previously defined during histological annotation. Many of the studies performed to date have concerned tumors with welldelineated borders. Heterogeneity complicates the workflow, from sample preparation (one must ensure all tissue types in the tumors provide good quality MSI data) to histological annotation and selection of comparable areas across all patient samples. Below we describe how we optimized a MALDI MSI protocol for analyzing proteins for fresh frozen high-grade sarcoma tissues Cryo-sectioning & randomization A semi-supervised block randomization was used to distribute the patient tissue sections between and within slides in order to minimize any potential sources of bias during MSI data acquisition (pseudo-code available in Carreira et al. 107 ). For MALDI MSI 12 µm thick sections were cut at -20 C in a cryostat and thaw mounted onto poly-l-lysine coated indium-tin-oxide glass slides (Bruker Daltonik, Bremen, Germany). The sections were stored at -80 C until use. After MALDI MSI data acquisition the same tissue sections were H&E stained and histologically annotated. Although tissue thickness in sample preparation for MALDI MSI experiment is not overly critical, μm thick sections are widely used owing to their greater robustness to the tissue preparation steps. Thinner sections provide superior histological information but can be very fragile; this is especially true for tissues containing bone such as osteosarcoma used in this study. Accordingly 12 µm thick sections were used for MALDI MSI and histological annotations; adjacent thin 5 µm sections were also cut for confirmation of histological annotation. 26

34 Chapter 2 Methodology Tissue washing A number of tissue washing procedures have been reported and compared 105,108 and which has also established that the wash that leads to the richest MSI datasets (i.e. number of molecular ions detected) can be tissue dependent 109. Bearing in mind the presence of numerous tissue/cell types in these complex histology tumors we compared three washing protocols to ensure rich molecular profiles could be obtained from all cell types (of the four complex histology sarcomas). A detailed description of the washing protocol is reported in Tab.2.1. For this comparison 0.5 μl of matrix solution (10mg/ml sinapinic acid in 60% Methanol: 0.1% TFA) was manually spotted (3-5 spots per tissue section) onto several high-grade leiomyosarcoma tissue sections. The mass spectra were then averaged, compared, and the protocol with the highest number of molecular ions, signal intensity and signal-to-noise ratio (S/N) was selected for the further experiments. Manual spotting of 0.5 μl of matrix solution was performed because while the mass spectra were similar to those obtained from tissue sections prepared with the ImagePrep matrix deposition system the preparation time was much less (1 min vs. 3-5 hrs for each tissue section). Accordingly multiple tissue wash protocols could be readily compared, using technical and biological repeats for statistical significance. Steps in washing protocol Time Ref. 70% ethanol 30 sec 100% ethanol 30 sec Carnoy s (60% ethanol, 30% chloroform and 10% acetic acid) 2 min 110 Carnoy s wash 100% ethanol 30 sec Milli-Q water 30 sec 100% ethanol 30 sec 500mM ammonium formate in water-acetonitrile solution (9:1 90 sec 111 v/v, 0.1% TFA, 0.1% Triton X-100) 2-step ethanol 70% ethanol 30 sec 112 wash 100% ethanol 30 sec 2-step 70% isopropanol 30 sec 105 isopropanol wash 95% isopropanol 30 sec Tab.2.1. Details of the three washing protocols compared here for the analysis of soft tissue and bone sarcomas. 27

35 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Matrix application The tissue sections were first washed, dried in a desiccator for 15 minutes, and then fiducial markers added using a Tipp-Ex pen. A matrix solution of 20 mg/ml sinapinic acid in 60% Methanol: 0.1% TFA was sprayed onto the tissue section using a pre-optimized method for the ImagePrep matrix deposition system (Bruker Daltonik, Bremen, Germany); the matrix deposition method was reported previously 102 and for convenience is reproduced here in Supplementary Tab.2.2. The matrix coated slides were then placed in the slide holder of the mass spectrometer and scanned in a flatbed scanner with 2400 dpi resolution (Epson V200 Photo). Note the resolution of the scanned image was much higher than that of the MSI data (100 µm pixel size corresponds to 254 dpi) to enable accurate coregistration with the histological image (see ) MSI data acquisition All MSI experiments were performed using an Ultraflextreme III MALDI-ToF/ToF mass spectrometer (Bruker Daltonik, Bremen, Germany) operated in linear mode, using a 100 µm pixel size, 500 laser shots per pixel (50 laser shots per position of a random walk within each pixel). Positively charged ions between m/z 2,000 and 23,000 were detected with a digitization rate of 1 GHz. All pixel mass spectra were processed with a smoothing algorithm (Gauss algorithm, width 2 m/z, 4 cycles) and a background subtraction (TopHat algorithm) during data acquisition using FlexAnalysis (version 3.4, Bruker Daltonik, Bremen, Germany). A preliminary experiment was performed to determine the combination of laser shots per pixel and spatial resolution that provided high quality mass spectra (rich mass spectra with high signal-to-noise). For this a histologically homogeneous high grade sarcoma sample was selected (to limit the impact of biological variance). Three consecutive tissue sections were prepared for MALDI MSI analysis and then divided into nine measurement regions, which were used to test different combinations of spatial resolution and laser shots per pixel: i) 150 µm; 300 shots; ii) 150 µm; 500 shots; iii) 150 µm; 1000 shots; iv) 100 µm; 300 shots; v) 100 µm; 500 shots; vi) 100 µm; 1000 shots; vii) 50 µm; 300 shots; viii) 50 µm; 500 shots; ix) 50 µm; 1000 shots. The optimization experiments were repeated on three consecutive days to ensure reproducibility. Mass spectral quality was evaluated by using ClinProTools

36 Chapter 2 Methodology (Bruker Daltonik, Bremen, Germany) to calculate i) the number of detected peaks (S/N >= 3); ii) average signal-to-noise ratio of detected peaks; iii) the number of non-excluded spectra per 100 spectra (spectra that were not alignable or null were excluded by ClinProTools inbuilt quality control metrics) Histological analysis After MSI data acquisition the slides were washed in 70% ethanol to remove any remaining matrix and then H&E stained. Ready-for-use solutions of hematoxylin and eosin were obtained from Klinipath BV (hematoxylin solution, article code ; eosin Y 1% alcohol solution, article code ) and used as provided (poured fresh as needed). A comparison between different staining times for the hematoxylin (1 min, 2 min, 3 min) and eosin (10 sec, 15 sec, 20 sec) staining was performed in order to achieve the best histological staining of the 12 µm thick MSI-analyzed tissue sections. High resolution digital images of the stained tissues were then recorded using a Pannoramic MIDI slide scanner (3DHISTECH Ld., Budapest, Hungary) Region-of-interest selection The high-grade sarcoma tissues were histologically heterogeneous, containing regions with different tumor grade (degree of differentiation). Tumor areas were categorized into well-differentiated, moderately differentiated and undifferentiated areas in order to reduce the contribution of cellular heterogeneity. For well-differentiated areas in OS osteoblastic areas were selected (AGHC). The moderately differentiated areas in UPS were less cellular and less pleomorphic but still lacked any line of differentiation at the histological or immunohistochemical level. The consecutive 5 µm thick sections were used for histopathological confirmation. Regions of interest in the MSI analyzed tissue sections were selected by expert pathologists (JVMG and AGHC) using the software Pannoramic Viewer (3DHISTECH Ld., Budapest, Hungary). Annotated histological images were imported into the MSI software FlexImaging 3.0 and co-registered with the MALDI MSI datasets using the fiducial markers present on each slide. The mass spectra from each ROI were then exported using FlexImaging 3.0 for data analysis. 29

37 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging MSI data analysis ClinProTools 3.0 was first used to assess MSI data quality; only datasets in which <40% of the spectra were null spectra or not alignable (2000 ppm maximum peak shift to at least 10% of the calibrant peaks) were used for subsequent statistical analysis. Furthermore, owing to the high variability of the MSI signals observed in the low mass range, m/z 2,000 3,000 was excluded from the statistical analysis. The criteria for patient and data inclusion are summarized in Tab Diagnostic biomarker analysis The spectra from each histologically-defined region-of-interest (ROI) were exported; 200 spectra were then randomly selected from each ROI and loaded into ClinProTools. Statistical comparisons of protein signatures from histologically comparable regions were performed using ANOVA, if the data were normally distributed, or the Kruskal-Wallis test, if the data were non-normally distributed. The Anderson-Darling test was used to determine the data s distribution (p 0.05 for a non-normal distribution; p > 0.05 for a normal distribution) and thus which test was appropriate. All p-values were adjusted for multiple hypotheses testing using the Benjamini & Hochberg correction. To limit any sampling bias (of the randomly selected spectra) all analyses were repeated five times. Protein signals were only considered significant if they shows significant discriminatory power in at least 80% of the five repetitions Prognostic biomarker analysis The exported spectra from each histologically-defined ROI were read into MATLAB R2011a (MathWorks, Natick, Massachusetts) using in-house developed routines for automated feature identification and extraction 113. The algorithm performs an alignment, total ion count normalization and removal of extreme pixels all of which are established methods for comparing the intensities of MALDI MSI peaks. The result is a project file containing the intensities of all detected peaks for each ROI. For these experiments, recording using a MALDI-TOF, peak area was used for intensity calculation. The resulting project file contained the intensities of 136 molecular ions from 105 annotated regions of well-differentiated, moderately differentiated and undifferentiated areas in MFS, LMS, OS and UPS. The association of MSI with clinical endpoints was then investigated in R (R 30

38 Chapter 2 Methodology Foundation for statistical Computing, Vienna, Austria). Two endpoints were considered:- overall survival - defined as the time period from date of surgery to the date of death or the last follow-up visit; metastasis-free survival - defined as the time period from the date of surgery to the date of appearance of metastasis or the last disease-free follow-up visit. The Significance Analysis of Microarrays tool in R ( samr package) was first used to identify all molecular ions that were statistically associated with overall/metastasis-free survival (q <= 0.05). Patients were then dichotomized based on the 3 rd quartile of intensity experience shows that molecular subgroups are usually found in 10% to 25% of the patients (e.g. HER2 overexpression 114, KRAS mutation 115 ) and their survival times were compared by the log rank test and Kaplan-Meier curves using the R survival package. Finally Cox proportional hazards regression model were used to assess any potential correlation of the protein signatures with known clinical co-factors, including tumor type and neoadjuvant chemotherapy. A molecular ion was considered significant if it had p-values <= 0.05 in both the log rank test and in the Cox model. A summary of the workflow for diagnostic and prognostic biomarker discovery is shown in Fig Fig Diagnostic and prognostic biomarker discovery workflows. Biomarkers for the differential diagnosis of sarcoma subtypes were identified using a univariate comparison 31

39 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging of peak intensities. Random subsets of MSI spectra were extracted from histologically specified region of each patient tissue and average spectra determined for each patient (repeated 5 times). The intensities of each protein ion were then compared between patients using statistical tests appropriate to the distribution of the data, and corrected for multiple testing. Prognostic biomarkers were identified by examining the association between protein ion intensities and patient survival. Peaks exhibiting a significant relationship with survival were first determined using significance analysis of microarrays (q 0.05, minimum median FDR). Patient prognosis was determined by comparing high and low expression groups in the Kaplan-Meier analysis. A Cox proportional hazards model as a multivariate analysis was used to assess the independence of the prognostic biomarkers to known clinical cofactors. Prognostic biomarkers were considered significant if the adjusted p value 0.05 in both Kaplan-Meier analysis and multivariate analysis Statistical classification of UPS For the classification of UPS samples, the average spectrum of each sample and region was exported from the MATLAB project file. These spectra from well-, moderately- and undifferentiated regions of LMS, MFS and OS were then loaded into the R statistical programming environment. To include other possible sources of UPS (i.e. not LMS, MFS or OS) a series of binary classifiers were created to demarcate e.g. LMS tumors from non-lms tumors. Each of these binary classifiers was trained using the well-differentiated regions of LMS, MFS and OS, and then validated using the undifferentiated regions present in the LMS, MFS and OS tissue samples. First, differentially diagnostic molecular ions between each pair of sarcoma type (except UPS) were determined similar as described in Fig.2.1, i.e. based on the Shapiro-Wilk normality test, a t- test or Wilcoxon test is used to determine significant differences in mass signal intensities. The resulting p-values were adjusted by Benjamini-Hochberg correction to reduced false-positive detections. These feature lists, containing only significantly discriminating protein signals for each pair of sarcoma type, were then merged to create three support vector machine binary classifiers to distinguish each specific tumor type (e.g. LMS) from the remaining types (e.g. OS and MFS). Validation on the undifferentiated regions provided the performance of the classifiers. Based on the achieved sensitivities, the cost parameter of the binary SVM classifiers was tuned: the costs for classifiers with high sensitivity were increased and for low sensitivity decreased. Afterwards, the binary classifiers were organized into a hierarchical dendrogram, ordered so that those with the highest 32

40 Chapter 2 Methodology specificity were applied first. This hierarchical binary classifier demarcated into 4 classes, LMS, MFS, OS and unknown, was validated using the undifferentiated regions present in the LMS, MFS and OS tissue samples, and finally applied to the UPS samples Molecular ion assignments Molecular ions that showed significant difference in intensity for distinguishing subtypes (diagnostic biomarkers) and predicting patient survival (prognostic biomarkers) were first assigned based on comparison with databases 64,65 that list molecular ions commonly detected by MALDI MSI, using a searching window of ±1000 ppm. 2.3 Results Here we report a detailed and successful workflow for diagnostic and prognostic biomarker discovery in high grade sarcomas using MALDI MSI, and describe how classifiers may be used to compare UPS (lacking any differentiation) with LMS, MFS and OS that do contain areas with cellular differentiation. The complete method is described, from sample preparation optimization, to biomarker discovery, to classification, to validation Washing protocol optimization Sample preparation is arguably the most crucial aspect that determines the success of a MALDI MSI experiment 106. We first compared three recently reported tissue washes to determine which gave the highest quality spectra and highest reproducibility from the heterogeneous soft tissue sarcoma tissues. Three consecutive tissue sections were obtained from the tumors and washed by the three different washing protocols: A) Carnoy's wash 110 ; B) 70% EtOH followed by 100% EtOH wash 112 ; C) 70% isopropanol followed by 95% isopropanol μl droplets of matrix solution were deposited throughout the tissue sections and MALDI mass spectra recorded. Fig.2.2 shows the average mass spectra obtained from the LMS tissues for each tissue wash, with indication of the number of peaks. It was found that the average mass spectrum from tissue sections 33

41 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging prepared with Carnoy s wash contained more peaks with S/N>3, and with equivalent or greater reproducibility Instrumental parameters Instrumental data acquisition parameters were then compared. Once the laser fluence was set, just above MALDI threshold, the mass spectral quality was compared as a function of spatial resolution and number of laser shots per pixel. Three consecutive sections of a LMS tissue sample were prepared using the Carnoy s wash, which had previously been found to lead to the greatest number of molecular ions, and then coated with sinapinic acid matrix. Nine measurement regions were defined on each tissue section, from which MSI data were acquired using different combinations of spatial resolution and number-of-laser-shots-perpixel. The spectra were then compared using ClinProTools. Fig.2.3 shows a comparison of the spectra by comparing the number of detected molecular ions (S/N>3), average S/N of each detected peak, and the number of null spectra (those that cannot be aligned or deemed of low quality by the ClinProTools software). The combination of 100 μm spatial resolution and 500 shots per pixel was selected, as it provided the best combination of number-of-protein-ions, average S/N of each detected peak, and the smallest proportion of null spectra. Fig.2.2 Comparison of tissue washing protocols. (a) Carnoy s solution based wash protocol, # peaks with S/N 3 were 35. (b) Two-step ethanol based wash, # peaks with S/N 3 were 29. (c) Two-step isopropanol based wash, # peaks with S/N 3 were 13 34

42 Chapter 2 Methodology Fig.2.3 Optimization of MALDI MSI data acquisition parameters. The number of peaks (red), average S/N of the detected peaks (blue) and proportion of alignable spectra (green) are plotted as a function of spatial resolution and number of laser shots per pixel Hematoxylin and eosin staining protocol Following MALDI MSI data acquisition any remaining matrix is removed and the tissue section stained with H&E 28. A high resolution optical image of the H&E stained tissue section, obtained using a digital slide scanner (here a 3D Histech Pannoramic MIDI), is then be registered to the MSI datasets. It is this combination of MSI and histology that enables the extraction of molecular profiles from histologically specified regions, from within the histologically heterogeneous sarcoma tissue sections. Tissue histology, and the quality of the staining, is thus central to the MALDI MSI experiment. The standard incubation times used in the pathology laboratory of Leiden University Medical Center (hematoxylin, 1 minute; eosin, 10 seconds) are optimal for 4-6 µm thick tissue sections that are cut for routine pathogical examination. To assess if this protocol was also optimal for the thicker 12 µm thick sections used for MALDI MSI (and following data acquisition) the incubation times for hematoxylin and eosin staining were investigated. Fig.2.4 shows a comparison of the histological images obtained for the different combinations of incubation times; it can be seen that the contrast of the hematoxylin staining (purple) was greatest for the 3-minute incubation, and that the 1 minute incubation of the standard protocol was not sufficient for the thicker 12 µm thick tissue sections. For the eosin staining (pink) the 20-second incubation led to excessive almost uniform staining. Accordingly it was decided to use the 3- minute hematoxylin and 15 second eosin incubations for the 12 µm thick tissue sections analyzed by MALDI-MSI. 35

43 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Fig.2.4 Optimization of hematoxylin and eosin staining of post MALDI MSI 12 μm thick tissue sections. Histological images are provided for different combinations of incubation times. 2mH = 2 minute hematoxylin incubation, 10sE = 10 second eosin incubation. Optimum was found to be 3 minutes of hematoxylin incubation and 15 seconds of eosin incubation Histological specification To compare the molecular signatures from histologically heterogeneous tumors it is essential that the tissue sections are first histologically annotated by experienced pathologists so that a scheme can be devised that details which histological regions can be compared in the different patient tissues. Once the histological criteria have been defined each patient s histological image is annotated to demarcate which regions will be used in the subsequent statistical analyses. We specified the tumor areas according to the degree of cellular differentiation (tumor grade) in order to i) reduce this source of variance during the subsequent statistical comparisons of MSI data, and ii) to investigate how the biomarkers may be influenced by the degree of cellular differentiation Diagnostic and prognostic biomarkers A full description of the diagnostic and prognostic biomarker experiments was recently reported 102. For this methodology description we provide a brief summary. The MALDI MSI data was extracted from the well-differentiated, 36

44 Chapter 2 Methodology moderately-differentiated and undifferentiated regions annotated in each patient tissue; for each region the mean average mass spectral signature was calculated and the intensities of the molecular ions compared between patients to determine candidate biomarker ions for differential diagnosis and prognosis. Candidate biomarker ions for differential diagnosis are those that exhibited a statistically significant difference in intensity between different tumor types; candidate biomarker ions for patient prognosis are those that exhibited a statistically significant difference in intensity between patient groups with short and long survival, and which were independent of other clinical co-factors such as neoadjuvant therapy. Twenty candidate biomarkers for differential diagnosis were obtained. However, the biomarkers for differential diagnosis were only statistically significant for the well-differentiated tumor areas, indicating that the loss of differentiation was accompanied by molecular changes that reduced the effectiveness of individual protein markers. The prognostic analysis found nine molecular ions that predicted patient outcome. It is important to assess if the prognostic biomarker ions exhibit any correlation with other clinical variables, such as tumor type, neoadjuvant therapy and adjuvant therapy, to ensure that the biomarkers are independent and not surrogates of other variables. Fig.2.5 shows Kaplan-Meier curves in which the patients have been divided according to the different clinical parameters and the associated p-values calculated from log-rank test between groups. Neoadjuvant chemotherapy: the Kaplan-Meier survival curves appeared different for patients that did or did not receive neoadjuvant therapy, with p-values for overall survival exhibiting a trend (p = 0.086) and for metastasis-free survival being statistically significant (p = 0.01). Tumor type: survival and metastasis-free survival didn t exhibit any statistically significant association with tumor type; that said MFS patients did appear a distinct group, exhibiting a better survival, for which the lack of statistical significance may reflect the small patient numbers for each distinct subtype. Adjuvant therapy: no significant differences in survival or metastasis-free survival were observed. Therefore, tumor type and neoadjuvant therapy status were included as co-factors in the multivariate Cox proportional hazards regression model. Only proteins that exhibited significant differences for both Kaplan-Meier and multivariate analysis were considered candidate prognostic biomarkers

45 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Fig.2.5. Survival analysis for patient follow-up data (no MSI related) for multivariate analysis co-factors inclusion. The influence of the available clinical data on overall survival (a-c) and metastasis-free survival (b-f) were compared. (a, d) show the Kaplan-Meier survival curves for patients that received neoadjuvant therapy and those that did not; (b, e) show the individual Kaplan-Meier survival curves for LMS, MFS, OS and UPS; (c, f) show the Kaplan-Meier survival curves for patients that received adjuvant therapy and those that did not. P values were calculated using a log-rank test. Based on the observed differences in survival, neoadjuvant therapy and tumor type were included in the multivariate assessment of biomarker independence Statistical classification of UPS UPS lacks any line of differentiation at the histological or immunohistochemical level. It is a diagnosis of exclusion for any sarcoma sample that cannot be more precisely categorized. UPS is genetically characterized by complex karyotype and is histologically pleomorphic and heterogeneous. The high-grade complex-karyotype tumors LMS, MFS and OS included regions of tissue characterized by welldifferentiated, moderately differentiated and undifferentiated cells. Accordingly it has been hypothesized that UPS represents a common end stage of such highgrade complex-karyotype tumors, e.g. high-grade leiomyosarcoma with loss of myogenic differentiation, high-grade osteosarcoma with loss of osteoid 38

46 Chapter 2 Methodology differentiation, and high-grade myxofibrosarcoma (MFS) with loss of myxoid areas, Fig.2.6a. We sought to investigate the relationships between the molecular signatures obtained from individual UPS patients using MALDI MSI, with those obtained from high grade LMS, MFS and OS. In order to better take into account the multivariate nature of MSI data we investigated if a classifier could be built that could relate UPS to LMS, MFS or OS. This endeavor was complicated by the realization that LMS, MFS and OS may not be the only tumors related to UPS, for example pleomorphic rhabdomyosarcoma with loss of rhabdoid differentiation may also be related to UPS. We accounted for this possibility by creating a series of binary classifiers, which when applied in a hierarchical setting could determine if UPS originated from LMS, MFS, OS or another unknown source, Fig.2.6b. First the welldifferentiated regions of the LMS, MFS and OS tumors were compared pairwise to determine which molecular ions statistically distinguished LMS from the rest (MFS and OS), MFS from the rest, and OS from the rest (Fig.2.7a). These feature sets were then used to build three SVM binary classifiers with linear kernel, which were trained on the well-differentiated regions and evaluated on the undifferentiated regions. Based on their performance, the parameter of constraints violation of the SVMs was tuned in which the cost for classifiers with a lower sensitivity was reduced and vice versa. This way, the classifier with the lowest sensitivity, LMS vs. REST, (50%) got a cost of 0.1, MFS vs. REST (60%) a cost of 1, and OS vs. REST (87.5%) the highest cost of These tuned classifiers were then organized in a hierarchy in which the classifier with the highest specificity (96.2%), LMS vs. REST, was applied first; followed by the classifier with the next highest specificity (91.7%), MFS vs. REST; followed by the final classifier, OS vs. REST (61.1%; Fig.2.7b). The hierarchical binary classifier was also validated on the undifferentiated regions of the LMS, MFS and OS tumor tissues, and resulted in an accuracy of 70.6 %, Fig.2.7c. The hierarchical classifier was then applied to the 13 MALDI MSI datasets of UPS patients. The regions highlighted by the classifier were then histologically reexamined with a specific remit to try to identify morphological features indicative of origin. Tab.2.3 shows the results of the hierarchical classification and morphological re-examination and demonstrates a high consensus of 85 %. 39

47 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Fig.2.6 (a) Origin of UPS as common end-stage of high grade complex genome sarcomas; (b) creation of hierarchical classifier in order to account for unknown origin of UPS. Fig.2.7 (a) Feature selection, determination of protein ions that differentially discriminate LMS, MFS and OS (only well differentiated regions used as previous results demonstrated these were the most discriminating); (b) organization of hierarchical binary classifier, and results of binary classifier training using well-differentiated regions of LMS, MFS and OS. The hierarchical classifier is organized with the most specific classifiers applied first; (c) results of the hierarchical binary classification of undifferentiated regions of LMS, MFS and OS. 40

48 Chapter 2 Methodology Patient # Hierarchical classifier Morphology examination 1 LMS 1 st LMS; 2 nd MFS; not OS 2 Unknown LMS 3 OS OS 4 OS OS 5 OS OS 6 OS OS 7 OS OS or MFS 8 MFS MFS focally 9 OS OS focally 10 OS OS 11 LMS LMS 12 MFS 1 st OS; 2 nd LMS; not MFS 13 LMS 1 st LMS; 2 nd MFS; not OS Consensus 85% Tab.2.3. Application of the hierarchical classifier to 13 UPS patient samples, and results of close morphologically re-examination of the tissues. 2.4 Discussion The four high grade complex karyotype sarcomas studied here, LMS, MFS, OS and UPS, are challenging for the pathologist because of possible histological overlap, and for the oncologist because of their sometimes only moderate or lack of response to chemo- and radiotherapy. OS is a bone sarcoma while the others are soft tissue sarcomas. The tumors are characterized by complex and heterogeneous histologies, which could (and was found) to be mirrored by complex and heterogeneous MALDI MSI datasets. Accordingly histology played a crucial role in the analysis of the MALDI MSI data. Thus following optimization of the MALDI MSI sample preparation (Fig.2.2) and the MALDI MSI instrument parameters for data acquisition (Fig.2.3) the histological staining was also optimized, and it was found that the standard staining method used in diagnostic pathology was not optimum for the thicker tissue sections analyzed by MALDI MSI (Fig.2.4). A large effort was placed into method optimization first because the effectiveness of any assay or biomarker discovery pipeline is intrinsically dependent on the MALDI MSI data quality. It is known that the optimum MALDI MSI sample preparation protocol can be sample (tissue) dependent 116,117. We compared three recently reported tissue washes and found that, for these sarcoma tissues, the Carnoy s fluid based washing protocol led to more molecular ion peaks and higher S/N than the ethanol or isopropanol based washing protocols. The strong dependence of the sample preparation protocol on the tissue is exemplified by 41

49 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging the observation that, for these sarcoma tissues, the Carnoy s fluid based tissue wash was superior to the ethanol and methanol based washes, whereas for atherosclerotic arteries the results were almost entirely contrary, i.e. isopropanol wash > ethanol wash > Carnoy s wash 109. The sarcoma and atherosclerosis experiments were performed in the same laboratory, using the same infrastructure, the same matrix deposition protocols and the datasets were recorded at similar times. To optimize MALDI MSI data quality the instrument parameters were also investigated. It was found that 100 µm spatial resolution and 500 shots per pixel provided spectra containing the largest number of molecular ions, greatest average S/N, and the smallest proportion of excluded spectra. At the smaller pixel size, 50µm, the number of peaks and mass spectral quality dropped dramatically with an increasing number of laser shots, and at the greatest number the tissue sections were damaged. This is consistent with a rapid ablation of the matrix at the smaller pixel sizes: increasing numbers of laser shots adds non-specific background to the mass spectra, lowering the mass spectral quality. It should be noted that developments in MALDI matrix deposition systems and laser irradiation technologies, for example the RapiFlex system from Bruker Daltonik, now enable protein signatures to be consistently obtained from pixel sizes as small as 20 um. Nevertheless matrix coating, laser fluence and number of laser shots need to be carefully set for high spatial resolution investigation. Two data quality filters were also used to ensure only high quality MALDI MSI data were used for the statistical analysis. Every pixel s mass spectrum, from every tissue section, was assessed using ClinProTools inbuilt quality control metrics. All mass spectra that contained few molecular ions or could not be aligned to the other mass spectra were excluded (data quality filter #1), and if a tissue section s MSI dataset had more than 40% excluded spectra it was excluded in its entirety (data quality filter #2). In MALDI MSI the large number of distinct molecular ions detected within each experiment makes the manual comparison of each MS image with the histological image entirely impractical, especially so if one also seeks to examine the correlations between the detected ions. Instead powerful statistical tools are used to interrogate the MALDI MSI datasets 118,119. In light of the histological heterogeneity of the sarcoma tissues we first defined the regions to compare on the basis of each tissue section s histology. The region-specific mass spectra were then extracted and the average mass 42

50 Chapter 2 Methodology spectrum calculated for each histological type patient (well differentiated, moderately differentiated and undifferentiated) from each patient. Statistical tests were then applied to determine which proteins were associated with tumor type (differential diagnosis biomarker) and with patient outcome (patient prognosis). A Cox proportional hazards model was used to test if the prognostic biomarkers were independent of known clinical variables. For the differential diagnosis biomarker discovery experiment it is essential that the original histopathological diagnosis of the patient tissues be of high confidence (i.e. that they represent histologically classical examples). It remains to be established whether the resulting biomarkers are equally applicable to the diagnostically more challenging cases. The presence of well-differentiated, moderately differentiated and undifferentiated regions within the same patient tissue were used to assess if the effectiveness of protein biomarkers were dependent on the tumor grade (degree of differentiation) and if classifiers could be developed to relate individual UPS patient tissues to high grade LMS, OS, or MFS. For this latter aspect the presence of well-differentiated and undifferentiated tissues in the same patient tissue is essential as the performance of a classifier can only be assessed if the ground truths of the samples are known. Twenty molecular ions differentiated between the well differentiated regions of LMS, MFS and OS 102 but did not differentiate the moderately or undifferentiated regions; however SVM based hierarchical classifiers were developed that could relate the undifferentiated regions with their neighboring well-differentiated regions with reasonable accuracy (Fig.2.6 and 2.7). When applied to the UPS patient series the classifier predictions were supported by close histological examination with 85% consensus (Tab.2.3). One of the other objectives of the sarcoma project was to identify common biomarkers for patient outcome, because a prognostic biomarker common to many/all sarcomas would enable patient stratification while circumventing many/any difficulties with definitive diagnosis. Common biomarkers are also beneficial from the logistical viewpoint of collecting a sufficient number of patient samples (the rarity of some subtypes may preclude subtype-specific investigations for example pleomorphic rhabdomyosarcoma was not included because insufficient patient samples were available). For clinical impact it is imperative that any prognostic biomarkers are independent and not surrogates for known clinical parameters; Fig.2.5 has been included here to explicitly demonstrate how to identify clinical variables that may be associated with the biomarker, and whose 43

51 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging relationship can then be assessed using established statistical tests (e.g. Cox Hazards model) Conclusions We have described in detail a successful biomarker discovery workflow for tumors characterized by highly heterogeneous histologies and MALDI MSI datasets. We describe how many aspects of the experiment need to be optimized prior to data acquisition of the patient series, and that a successful clinical experiment demands intimate interplay between histology and the MALDI MSI datasets. By focusing on specific histological regions (to reduce the molecular variation that is associated with different cell types) panels of proteins and classifiers were determined that could distinguish between different subtypes of sarcomas and predict patient survival. Classifiers were also built that show promise for establishing a relationship between UPS and high grade LMS, MFS and OS. 2.6 Acknowledgements The authors would like to acknowledge financial support from Q2 COMMIT, Cyttron II and the ZonMW Zenith project Imaging Mass Spectrometry-Based Molecular Histology: Differentiation and Characterization of Clinically Challenging Soft Tissue Sarcomas (No ). BB is funded by the Marie Curie Action of the European Union (SITH FP7-PEOPLE-2012-IEF No ). The authors also thank University Hospital Leuven for contributing an MFS patient sample. 44

52 Patient prognosis Differential diagnosis Chapter 2 Methodology Supplementary Information Supplementary information LMS MFS UPS OS # patients Gender (M/F) 5 / 6 4 / 8 (1 unknown) 8 / 3 11 / 3 Median age [years] # patients Gender (M / F) 6 / 5 3 / 7 7 / 4 9 / 4 Median age [years] Neoadjuvant therapy 2 / 11 0 / 10 2 / 11 8 / 13 Adjuvant therapy 7 / 11 6 / 10 7 / 11 9 / 13 Median survival [months] 25 64% survival at 103 months Median metastasis-free 73% metastasis-free survival 10 survival [months] at 52 months 10 9 Supplementary Tab.2.1. Clinical characteristics of the patient series. Reproduced with permission from Lou et al 102. Note: patient inclusion criteria were defined separately for the differential diagnosis biomarker discovery and patient prognosis biomarker discovery experiments (Tab.2.2 of main manuscript) and so patient numbers differed for each analysis type and from the total sample number. 45

53 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Spray phase Variables Value matrix thickness 0.1 V +/- 0.0 V cycles 2-3 Initialization_1 spray power 20% +/- 10% fix time 1.50 sec incubation time 5 sec +/- 0.0 sec dry time 20 sec matrix thickness 0.3 V +/- 0.0 V cycles 2-8 Initialization_2 spray power 20% +/- 10% fix time 2.50 sec incubation time 15 sec +/- 0.0 sec dry time 45 sec Phase 2 dry time 60 sec matrix thickness 0.2 V +/- 0.0 V cycles 2-22 spray power 17% +/- 10% Phase 3 spray time 0.1 V (sensor controlled) incubation time 20 sec +/- 30 sec dry (sensor control) complete dry every cycle safety dry 15 sec matrix thickness 0.3 V +/ V cycles spray power 20% +/- 10% Phase 4 spray time 0.2 V (sensor controlled) incubation time 20 sec +/- 30 sec dry (sensor control) 30% complete dry every 2nd cycle safety dry 10 sec matrix thickness 0.25 V +/ V cycles 4-15 spray power 20% +/- 10% Phase 5 spray time 0.3 V (sensor controlled) incubation time 20 sec +/- 30 sec dry (sensor control) 40% complete dry every 3rd cycle safety dry 10 sec matrix thickness 0.25 V +/-20 V cycles 4-15 spray power 20% +/- 10% Phase 6 spray time 0.3 V (sensor controlled) incubation time 20 sec +/- 30 sec dry (sensor control) 40% complete dry every 5th cycle safety dry 10 sec Supplementary Tab.2.2. Matrix deposition method for ImagePrep device. Reproduced with permission from Lou et al

54 Chapter 3 High-grade sarcoma diagnosis and prognosis: Biomarker discovery by mass spectrometry imaging. Lou S, Balluff B, de Graaff MA, Cleven AH, Briaire-de Bruijn I, Bovée JV, McDonnell LA. Proteomics Jun;16(11-12): doi: /pmic

55 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Abstract The combination of high heterogeneity, both intra-tumoral and inter-tumoral, with their rarity has made diagnosis, prognosis of high grade sarcomas difficult. There is an urgent need for more objective molecular biomarkers, to differentiate between the many different subtypes, and to also provide new treatment targets. Mass spectrometry imaging (MSI) has amply demonstrated its ability to identify potential new markers for patient diagnosis, survival, metastasis and response to therapy in cancer research. In this study we investigated the ability of MALDI-MSI of proteins to distinguish between high grade osteosarcoma (OS), leiomyosarcoma (LMS), myxofibrosarcoma (MFS) and undifferentiated pleomorphic sarcoma (UPS) (Ntotal = 53). We also investigated if there are individual proteins or protein signatures that are statistically associated with patient survival. Twenty diagnostic protein signals were found characteristic for specific tumors (p <= 0.05), amongst them acyl-coa-binding protein (m/z 11,162), macrophage migration inhibitory factor (m/z 12,350), thioredoxin (m/z 11,608) and galectin-1 (m/z 14,633) were assigned. Another nine protein signals were found to be associated with overall survival (p <= 0.05), including proteasome activator complex subunit 1 (m/z 9,753), indicative for non-os patients with poor survival; and two histone H4 variants (m/z 11,314 and 11,355), indicative of poor survival for LMS patients. Keywords biomarker discovery; intratumor heterogeneity; mass spectrometry imaging; soft tissue sarcoma Abbreviations MSI Mass spectrometry imaging; OS Osteosarcoma; LMS Leiomyosarcoma; MFS Myxofibrosarcoma; UPS Undifferentiated pleomorphic sarcoma; STS Soft tissue sarcoma 48

56 Chapter 3 Diagnostic and Prognostic Proteins 3. High Grade Sarcoma Diagnosis and Prognosis: Biomarker Discovery by Mass Spectrometry Imaging 3.1 Introduction High grade sarcomas can be diagnostically challenging tumors because of their rarity, comprising just 1% of all malignant tumors 1 and the fact that over fifty histological subtypes are described with overlapping morphology 3. This combination has made their correct diagnosis and treatment difficult, because, unless based in dedicated centers of expertise, a pathologist is unlikely to encounter many of the subtypes. Additional complications arise from the intratumor heterogeneity known to be present in these tumors 4 ; histologically identical but molecularly distinct locations may differ in clinical behavior and/or require different treatment. High grade sarcomas are a heterogeneous group of sarcomas, often lacking clearly defining features at light microscopy that may indicate a line of differentiation. For instance, osteoid deposition, defining osteosarcoma, or myxoid areas, defining myxofibrosarcoma, or pleomorphic lipoblasts, defining pleomorphic liposarcoma, may be absent in small biopsy specimens. Immunohistochemistry is used to exclude metaplastic carcinoma, leiomyosarcoma, pleomorphic rhabdomyosarcoma, or angiosarcoma, but expression may be heterogeneous. If the presently available techniques do not identify any line of differentiation, the tumor is labelled undifferentiated pleomorphic (or spindle cell) sarcoma 1. Though heterogeneous, all of these tumors are high grade sarcomas, characterized by complex genomes and often having a poor outcome. For soft tissue sarcomas the mainstay of treatment is surgical resection, as the most effective adjuvant treatments only achieve response rates of approximately 20% 120. Response rates are highly variable between different sarcoma subtypes emphasizing the critical role of an accurate diagnosis. However full concordance between primary diagnosis and second opinion has been determined to be just 56%, and more than 40% were modified at the second reading 5. Accurate diagnosis can influence outcome and treatment because different histological subtypes have different clinical behavior and require different treatment 100. There is thus an urgent need for novel diagnostic markers, to better discriminate between these poorly differentiated tumors, and for prognostic biomarkers, to predict outcome, especially on small needle biopsies. MALDI-MS has been successfully used for the direct molecular assessment of 49

57 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging tissue samples. The analysis is label-free, no a-priori knowledge is needed, and it is able to simultaneously measure hundreds of biomolecular ions. Spatiallycorrelated analysis, mass spectrometry imaging (MSI), can reveal how each of these biomolecular ions varies in tissue sections 85. Following data acquisition, a histological image of the same tissue section can be recorded and registered to the MSI dataset. This seamless integration of MSI and histology enables the mass spectral signatures of distinct cells and groups of cells to be acquired within their correct histological context. There is growing evidence that MSI is having an impact in disease detection, particularly cancer 25,86. The differential MS profiles found in tumors can be used to identify candidate biomarkers 43,64, and when combined with clinical outcomes identify MS signatures associated with prognosis 51 or response to therapy 52. MSI has been previously used to investigate the molecular make-up of sarcomas. In 2006 Caldwell et al. 80 reported that the cancer-associated proteins in undifferentiated pleomorphic sarcoma (UPS, previously malignant fibrous histiocytoma) extend into the surrounding morphologically healthy tissue. The ability of MSI to uncover molecular variation that does not mirror histology was further demonstrated in two subsequent studies of myxofibrosarcoma 4,73, in which an apparent clonal structure was evident in the molecular maps but not present in the histological images. Willems et al. 4 also demonstrated that MSI could find molecular markers to aid the differential diagnosis of myxoid sarcomas, reporting tissue type and tissue grade specific protein biomarkers for differentiating between myxofibrosarcoma and myxoid liposarcoma. The small protein cysteinerich intestinal protein 1, identified by MSI as a prognostic marker in gastric cancer 51 and breast cancer 82, was subsequently demonstrated to be associated with a favorable outcome and a lower metastatic rate in osteosarcoma patients 81. MSI has also been applied to formalin-fixed paraffin embedded (FFPE) tissues, including tissue microarrays (TMA s) containing the needle biopsies of large cohorts of patients. Groseclose et al. were the first to report MSI analysis of TMA s 36 ; using a TMA containing 112 lung cancer patient biopsies it was shown that MSI could identify protein biomarkers that distinguish biopsies from adenocarcinoma from squamous cell carcinoma biopsies. The ability to rapidly analyze large groups of patients in such TMA s has led to the identification of phenotypic biomarkers in, amongst others, oesophageal cancer 121, prostate cancer 122, and bladder cancer 123. These examples, as well as others from different cancers, indicate that MSI has great potential for providing novel abilities for 50

58 Chapter 3 Diagnostic and Prognostic Proteins improved differential diagnosis and patient management. In this study we investigated the ability of MALDI-MSI to distinguish between the high grade sarcomas osteosarcoma, leiomyosarcoma, myxofibrosarcoma and UPS, and if there are individual proteins or protein signatures that are statistically associated with diagnosis, or patient survival and the development of metastasis. Owing to the morphological heterogeneity often seen in these tumors we also assessed how the biomarkers were influenced by tumor grade/degree-of-cellulardifferentiation. Two distinct approaches have been used: the first follows the established clinical MSI protocol of using virtual-microdissection of the aligned histological image to extract each patient s average tumor-specific mass spectral profile. The second approach incorporates intratumor heterogeneity into the discovery pipeline 79 and so is able to identify biomarkers even if they exhibit a high degree of heterogeneity in the tumor samples Materials and methods Tissue specimens and sample cohorts Fresh frozen tumor samples with diagnosis of high grade myxofibrosarcoma (MFS), leiomyosarcoma (LMS), undifferentiated pleomorphic sarcomas (UPS) and osteosarcoma (OS) were collected from the archive of the department of Pathology of Leiden University Medical Center, Leiden, the Netherlands, except one MFS sample which was obtained from University Hospital Leuven, Leuven, Belgium. Slides were re-evaluated histologically and classified according to the 2013 World Health Organization criteria. MFS and LMS cases were histologically graded according to the La Fédération Nationale des Centres de Lutte Contre le Cancer. Only tissue blocks containing >60% viable tumor material were selected. The final patient series of tumors comprised 13 patient samples of MFS, 12 patient samples of LMS, 12 patient samples of UPS and 16 patient samples of OS. Tab.3.1 shows the clinicopathological data of the patient series. All specimens were handled according to the ethical guidelines described in Code for Proper Secondary Use of Human Tissue in the Netherlands of the Dutch Federation of Medical Scientific Societies. 51

59 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Tissue preparation 12 µm thick tissue sections were cut at -20 C in a cryostat and thaw mounted onto poly-l-lysine coated indium-tin-oxide glass slides (Bruker Daltonics, Bremen, Germany). The sections were stored at -80 C until use. All the tissues were washed prior to matrix deposition as follows: i) 30 seconds in 70% ethanol; ii) 30 seconds in 100% ethanol; iii) 2 minutes in Carnoy s solution (60% ethanol, 30% chloroform and 10% acetic acid); iv) 30 seconds in 100% ethanol; v) 30 seconds in Milli-Q water; vi) 30 seconds in 100% ethanol; vii) 90 seconds in 500mM ammonium formate in water-acetonitrile solution (9:1; v/v, 0.1% TFA, 0.1% Triton X-100). All solvents were chilled before use. The washed tissue sections were dried in a desiccator for 15 minutes then fiducial markers were added near to the tissue using a Tipp-Ex pen. Diagnostic biomarker analysis Prognostic biomarker analysis 52 No. of patients Male vs. Gender Female Median Age [years] No. of patients Male vs. Gender Female Median Age [years] Neoadjuva nt Therapy treatment Adjuvant treatment Median overall survival Length of [months] followup metastasis- Median free survival [months] Undifferentiated Leiomyosarco Myxofibrosarcoma pleomorphic Osteosarcoma ma sarcoma vs. 6 4 vs. 8 (1 not available) 8 vs vs vs. 5 3 vs. 7 7 vs. 4 9 vs / 11 0 / 10 2 / 11 8 / 13 7 /11 6 / 10 7 / 11 9 / % survival probability at max. follow-up time (103 months) 73% metastasisfree probability at max. follow-up time (52 months) Tab.3.1. Clinicalpathological characteristics of the patient series

60 Chapter 3 Diagnostic and Prognostic Proteins The matrix solution, 20 mg/ml sinapinic acid in 60% methanol: 0.1% TFA was sprayed onto the tissue sections using an ImagePrep matrix deposition device (Bruker Daltonics, Bremen, Germany) (spraying method reported in Chapter 2 Supplementary Tab.2.2). Prior to MSI data acquisition the matrix coated slides were placed in the MSI slide holder and 2400dpi resolution images recorded using a flatbed scanner (Epson V200 Photo) Data acquisition A semi-supervised block randomization was used to distribute the patient tissue sections between and within slides in order to minimize any potential sources of bias during MSI data acquisition (pseudo-code available in Carreira et al. 107 ). All MSI experiments were performed using an Ultraflextreme III MALDI-ToF mass spectrometer (Bruker Daltonics, Bremen, Germany), 100 µm pixel size, 500 laser shots per pixel (50 laser shots per position of a random walk within each pixel). Positively charged ions between m/z 2,000 and 23,000 were detected with a digitization rate of 1 GHz. All pixel mass spectra were processed with a smoothing algorithm (Gauss algorithm, width 2 m/z, 4 cycles) and a background subtraction (TopHat algorithm) during data acquisition using FlexAnalysis (version 3.4, Bruker Daltonics, Bremen, Germany). After MSI data acquisition the slides were washed in 70% ethanol to remove any remaining matrix and then H&E stained (protocol reported in Supplementary Tab.3.1). High resolution digital images of the stained tissues were then recorded using a Pannoramic MIDI slide scanner (3DHISTECH Ld., Budapest, Hungary) and co-registered to the MSI datasets in FlexImaging 3.0 (Bruker Daltonics, Bremen, Germany) Data analysis The stained tissue sections were histopathologically annotated by expert pathologists (M.A.G and J.V.M.G.B.). In order to reduce the influence of the tumors cellular heterogeneity on the MSI results, tumor areas were annotated in all samples according to their differentiation grade, namely well-, moderately, and un-differentiated, and using a consistent annotation scheme for the different STS subtypes (moderately differentiated tumor areas in UPS were less cellular and less pleomorphic, although still lacking any line of differentiation at the histological or 53

61 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging immunohistochemical level). For the well differentiated areas in OS we specifically focused on osteoblastic areas (selected by A.H.G.C.). Fig.3.1 shows a histological overview of the annotations used in the analysis. The quality of all MSI datasets was first evaluated in ClinProTools 3.0 (Bruker Daltonics, Bremen, Germany). Protein profiles from histologically comparable tumor areas were extracted and processed in ClinProTools 3.0 (settings reported in Supplementary Tab.3.2). Any dataset in which 40% of the spectra were excluded by ClinProTools inbuilt quality control metrics (not alignable to other spectra or null spectra) was excluded from further analysis. Owing to the high variability of the MSI signals observed in the low mass range, region of m/z 2,000 3,000 was excluded from the statistical analysis. Tumor area Annotation examples Osteosarcoma (OS) Leiomyosarcoma (LMS) Myxofibrosarcoma (MFS) Well differentiated/ Moderately differentiated/ Undifferentiated Undifferentiated pleomorphic sarcoma (UPS) Moderately differentiated/ Undifferentiated Fig.3.1 Overview of histological annotations Diagnostic biomarker analysis After mass spectral processing in ClinProTools3.0 (settings shown in Supplementary Tab.3.2), statistical comparisons of protein signatures from histologically comparable regions were performed using ANOVA, if the data were normally distributed, or the Kruskal-Wallis test, if the data were non-normally distributed. The Anderson-Darling test was used to determine the data s distribution (p 0.05 for a non-normal distribution; p > 0.05 for a normal distribution) and thus which test was appropriate. All p-values were adjusted for 54

62 Chapter 3 Diagnostic and Prognostic Proteins multiple hypotheses testing using the Benjamini & Hochberg correction Prognostic biomarker analysis The spectra from annotated regions were read into MATLAB R2011a (MathWorks, Natick, Massachusetts), which also included the bioinformatics and image processing toolboxes. The spectra were loaded, normalized to their total-ioncount, and then reduced using an automated feature identification and extraction routine (full settings shown in Supplementary Tab.3.3), using a modified form of the algorithm previously reported by our group 113. Briefly, basepeak mass spectra were created from each region of each tissue section. The basepeak spectrum displays the maximum intensity of every channel of the aligned mass spectra and is more effective for detecting peaks with localized expressions 113. The basepeak mass spectra from the different patient samples and different regions were then aligned on common peaks (peaks present in at least 85% of all samples). The global basepeak mass spectrum was then calculated (Supplementary Fig.3.1) and the protein peaks detected using an adapted version of the LIMPIC package 124. The resulting peak list was then used to extract the peak intensities from the annotated regions in each patient s MSI dataset; this reduced and more computationally-manageable MSI data was then placed into a project-specific data cube using spatial offsets to separate the different patient samples. The project data cube contains the MSI data of all samples, with the corresponding mass spectral data in the z-dimension. The association of MSI protein signals from each tumor type with clinical endpoints was then investigated in R (R Foundation for statistical Computing, Vienna, Austria). All protein peaks that exhibited a significant relationship between peak intensity and time data (overall survival time, which was defined as the time period from date of surgery to the date of death or the last follow-up visit; and metastasis-free survival time, which was defined as the time period from the date of surgery to the date of appearance of metastasis or the last disease-free follow-up visit) were first identified using the significance analysis of microarrays package in R (q <=0.05; minimum median FDR). Patients were then dichotomized based on the 3rd quartile of intensity and their time data again compared with Kaplan-Meier analysis using the survival package. All resulting significant masses were then evaluated using multivariate analysis (Cox proportional hazards regression model) to assess any correlation of these masses to clinically relevant co-factors such as tumor type and neoadjuvant 55

63 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging therapy Biomarker assignments Protein ions found to exhibit statistically significant differences for patient diagnosis or prognosis were assigned based on comparison with databases 64,65 of protein ions commonly detected by MALDI-MSI using a searching window of ±1000 ppm Immunohistochemical validation Selected protein biomarkers were validated using immunohistochemistry on formalin fixed paraffin embedded tissues. Supplementary Tab.3.4 provides a detailed description of the immunostaining protocol Intratumor heterogeneity investigation For investigating molecular intratumor heterogeneity, we first performed virtual microdissection to isolate the regions representing different differentiation grades (well-, moderately-, and un-differentiated tumor areas). The association of the mass spectral signatures from each differentiation region with clinical endpoints was investigated using the methods described above for prognostic biomarker analysis. Secondly, morphologically invisible subpopulations within specific differentiation regions were investigated using the method reported by Balluff et al. 79. Briefly, a segmentation algorithm based on the consensus of five multivariate methods was first used to delineate subpopulations with distinctive mass spectral profiles 73. The molecular segmentation was run with the number of expected tumor subpopulations/clusters (k) ranging from 2 to 10. The clinical importance of each cluster was then determined. Patients were grouped based on the appearance of the clusters in each patient s MSI dataset, using a minimum threshold defined as the contribution that may be expected on chance alone (i.e. 1/k, where k is the number of clusters). The resulting clusters were then analyzed using the methods described above for prognostic biomarker analysis. 3.3 Results Here we investigated whether MALDI-MSI could aid the differential diagnosis of 56

64 Chapter 3 Diagnostic and Prognostic Proteins the more common high grade sarcoma subtypes (including LMS, MFS, OS, and UPS) and to identify prognostic biomarkers for survival and metastasis. Owing to the high morphological complexity of high grade sarcomas the first step of the analysis was a histopathological annotation of tumor areas with well, moderate and un-differentiated regions. Spectra were then extracted from these regions by virtual microdissection to investigate how biomarker performance is affected by tumor grade /degree of cellular differentiation Identification of diagnostic biomarkers Using the software ClinProTools a set of 200 mass spectra were randomly selected from each patient s microdissected MSI dataset. Protein signals (m/z values) with significant discriminatory power to distinguish between the different high grade sarcoma subtypes were identified as described in the Materials and methods section. To exclude any pixel-selection related bias on the results, the procedure of random pixel/mass spectrum selection, processing and statistical testing was repeated five times. Protein signals were considered potentially diagnostically relevant if they showed significant (p <= 0.05) intensity differences specific to a single tumor type in 80% of the runs. In this way, twenty protein signals from well-differentiated regions of LMS (n=4), MFS (n=9) and OS (n=9) were found that exhibited a statistically significant difference in signal intensities, Supplementary Tab.3.5. By comparison with published databases of proteins commonly detected by MALDI-MSI 64,65, four of these could be assigned as acyl-coa-binding protein (m/z 11,162), thioredoxin (m/z 11,608), macrophage migration inhibitory factor (m/z 12,350) and galectin-1 (m/z 14,633). Tab.3.2 lists the biological function and the cancer context of these proteins, and Fig.3.2 gives example of the differential expression of m/z 6,226 and m/z 11,608 (assigned as Thioredoxin) between LMS, OS, and MFS tumors. 57

65 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging m/z observe d 11,608 12,350 14,633 11,162 Indication High expressed in leiomyosarcoma; Low expressed in myxofibrosarcom a High expressed in leiomyosarcoma; Low expressed in myxofibrosarcom a High expressed in leiomyosarcoma; Low expressed in myxofibrosarcom a High expressed in leiomyosarcoma; Low expressed in myxofibrosarcom a Protein-ID UniProt -ID Thioredoxin 12 5 P10599 Macrophage migration inhibitory factor 15 Galectin Acyl-CoAbinding protein 53 P14174 P09382 P07108 Protein function Redox reaction participator Inflammator y response related Regulators of immune responses Involved in the metabolism of fatty acids Tumor relevance Highly expressed in lung carcinomas 12 6 Diagnostic marker to discriminate between subtypes of STSs 127 Tumor markers of renal cell carcinoma 129 Stearoyl-CoA Desaturase 1: diagnostic marker to discriminate between subtypes of STSs 127 Tab.3.2 Assignments of diagnostic biomarkers, including information regarding protein function and tumor relevance. 58

66 Chapter 3 Diagnostic and Prognostic Proteins Fig.3.2 Visualization of diagnostic biomarkers. Diagnostic biomarkers were found characteristic of well-differentiated tumor areas of high grade sarcomas. Visualization of two examples, m/z 11,608 and m/z 6,226, are shown for representative samples of LMS, OS and MFS. Histological images are shown in (a-c), (d-f) depict the differential expressions of m/z 11,608 (thioredoxin) and Fig.3.2 (g-i) depict the differential expressions of m/z 6,226. (j) compares the average spectra from LMS, OS and MFS, and highlights the discriminating protein signals at m/z 11,608 and m/z 6, Identification of prognostic biomarkers Prognostic protein signals (associated with metastasis or survival) were searched for within each high grade sarcoma subtype (MFS, LMS, OS, UPS) and for a soft tissue sarcoma set consisting of all except OS (bone sarcoma) samples (termed non-os subset). After collective mass spectral processing of all samples in MATLAB, the statistical analysis was done separately for each subsets as described in 59

67 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Materials and methods section. In total, nine protein signals showed significant differences in overall survival (Supplementary Tab.3.6) after multivariate analysis evaluation (p <= 0.05). Fig.3.3 (a-d) show the corresponding statistical plots of four of the prognostic protein signals. By comparison with published databases of proteins commonly detected by MALDI-MSI 64,65, four of the nine could be assigned as proteasome activator complex subunit 1 (PSME1, m/z 9,753, indicative for non- OS patients with poor survival), two histone H4 variants (m/z 11,314 and m/z 11,355, indicative for LMS patients with poor survival) and hemoglobin subunit beta (m/z 15,877, indicative for MFS patients with poor survival). Fig.3.3 Kaplan-Meier survival plots of prognostic biomarkers. Examples of survival associated protein ions in soft tissue sarcomas (non-os) and undifferentiated pleomorphic sarcoma (UPS) patients are shown in (a-d), in which a higher expression of the protein are associated with poor survival. (e) shows a protein signal (m/z 6,653, cytochrome c oxidase subunit 2) that was found only when the analysis was limited to undifferentiated areas, and was indicative of poor survival for leiomyosarcoma (LMS) patients. m/z 8,093, whose prognostic value is shown in (f) gives a trend for indicating metastasis-free survival in UPS patients (with p value of 0.03 in Kaplan-Meier analysis while p value of 0.06 in multivariate analysis). All of these proteins have previously been detected by MALDI-MSI and associated with cancer including distinct aspects linked to its progression, such as 60

68 Chapter 3 Diagnostic and Prognostic Proteins microvascular invasion 130 and stromal activation 131 (Tab.3.3). Fig.3.4 shows the differential expression of proteasome activator complex subunit 1 (m/z 9,753) in two UPS patients with a different survival time and its overall prognostic value in non-os patients. The expression of PSME1 was also validated using immunohistochemistry. Formalin-fixed paraffin-embedded tissues from the same patients whose fresh-frozen tissues were analyzed by MSI were subject to immunohistochemistry. The same pattern of expression that was observed with MSI, namely higher expression from poorer prognosis patients, was also observed by immunohistochemistry (see Supplementary Fig.3.2). m/z observed 9,753 11,314 11,355 15,877 Indication Poor survival in soft tissue sarcoma patients Poor survival in leiomyosarcom a patients Poor survival in myxofibrosarc oma patients Protein-ID Proteasome activator complex subunit 1 39 Modified histone H4 132 Modified histone H4 132 Hemoglobin subunit beta 128 UniProt -ID Q P62805 P62805 P68871 Protein function Implicated in immunoproteas ome assembly and required for efficient antigen processing Transcription regulation Involved in oxygen transport Tumor relevance Marker of stromal activation 131 Microvascul ar invasion in hepatocellul ar carcinoma 130 Potential serum biomarkers of ovarian cancer 133 Tab.3.3 Assignments of prognostic biomarkers, including information regarding protein function and tumor relevance. m/z 11,314 is assigned as modified histone H4 with 1 methionine loss, 1 acetylation and 2 methylations; m/z 11,355 is assigned as modified histone H4 with 1 methionine loss, 2 acetylations and 2 methylations. 61

69 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Fig.3.4 m/z 9,753 as prognostic biomarker. (a) shows the Kaplan-Meier survival plot for m/z 9,753 (proteasome activator complex subunit 1, or PSME1) for soft tissue sarcoma patients (non-os subset) with good (blue curve) and poor (red curve) survival. (b) presents a comparison of the average spectra of two undifferentiated pleomorphic sarcoma (UPS) samples, one with better survival (blue mass spectrum) and one with poorer survival (red mass spectrum). Histological images and MSI visualizations for these two samples are shown in (c-f) Identification of phenotypic intratumor heterogeneity To investigate whether distinct subpopulations within the tumors were responsible for the clinical correlation, the prognostic biomarker analysis was repeated using two different approaches (Supplementary Fig.3.3). This was first performed by using the data from distinct regions, defined by their level of cellular differentiation (well-, moderately, or un-differentiated regions). This approach resulted in the identification of two additional prognostic proteinsm/z 6,281 and m/z 6,653 (Supplementary Tab.3.7) compared to the prognostic biomarkers identified previously in which the average mass spectrum from the entire MSI dataset was tested. Both were indicative for LMS patients (by using the data from undifferentiated tumor areas) with poor survival and protein ion detected at m/z 6,653 could be assigned as cytochrome C oxidase subunit 2, Fig.3.3(e) shows the Kaplan-Meier statistical plot. 62

70 Chapter 3 Diagnostic and Prognostic Proteins We then sought to investigate if there were any tumor subpopulations that determined patient prognosis within a same differentiation grade - the histologically homogeneous tumor areas. We have previously demonstrated how MSI is able to uncover tumor subpopulations in histologically identical regions of tissue 4,73, and recently shown that when the patients clinical data is available these tumor subpopulations may be statistically associated with patient survival and metastasis 79. Here we applied these routines to the MSI data from each STS tumor type and differentiation grade. For example, Fig.3.5 shows the results of this analysis for moderately differentiated tumor areas in OS. First, virtual microdissection was performed to isolate the MSI data from the moderately differentiated tumor areas in OS tumor samples (Fig.3.5a, 3.5b). The mass spectra from the isolated areas of all OS tumor samples were then analyzed by the agreement analysis (a consensus plot of five independent multivariate data analysis techniques); distinct tumor subpopulations were revealed that were characterized by distinct mass spectral profiles. Fig.3.5c shows that the moderately differentiated areas of this patient s primary tumor were thus demarcated into mainly two subpopulations that were histologically identical (Fig.3.5e and 3.5f). By comparing the presence of tumor subpopulations with the clinical data it was found that these subpopulations were statistically associated with different overall survival (Fig.3.5d). 63

71 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Fig.3.5 Analysis of tumor subpopulations. (a) shows an example of the OS patient samples analyzed by MALDI-MSI and subsequently histopathologically annotated; (b) shows the moderately differentiated tumor areas after virtual microdissection. This process was repeated for all moderately differentiated regions of OS patients and then the pixelassociated spectra clustered on the basis of their mass spectral profiles; (c) shows two molecularly-distinct tumor subpopulations in purple (cluster 4) and in pink (cluster 5) revealed by pixel-associated spectra clustering on the moderately differentiated areas of the OS patient tissue shown in (b); (d) shows these two clusters were statistically associated with different overall survival by linking the presence of the clusters from all patient samples to the available clinical data (OS patients with moderately differentiated areas); (e-f) provide histologically equivalent tumor areas in magnifications. 3.4 Discussion High grade sarcomas can be difficult to classify, as their morphology is often heterogeneous, as demonstrated in Fig.3.1. Better differentiated areas, essential for a correct diagnosis, may be lacking, especially in small biopsy specimens. To reduce the impact of the histological variability on the analysis, which would contribute a significant source of variance, a virtual microdissection for tumor areas was performed to isolate the MSI data from morphologically distinct regions. These tumor areas were also annotated according to their cellular differentiation 64

72 Chapter 3 Diagnostic and Prognostic Proteins grades to investigate if biomarker performance is affected by the degree of cellular differentiation. Our diagnostic biomarkers were only detected from welldifferentiated tumor areas, indicating that the loss of the tumor cell s distinctive character was accompanied by the loss of the effectiveness of the differentially expressed markers. Previous studies using gene expression profiling have reported differentially expressed genes as well as broader similarities between the complex genome tumors investigated here. For example, Villacis et al. 9 reported 587 genes that are differentially expressed between LMS and UPS, but did not detect a clear separation using hierarchical clustering. Several of the differentially expressed genes were validated using quantitative PCR and the most discriminating, SRC, confirmed at the protein level with IHC. An earlier study reported similar findings, namely that a fraction of the analyzed genes were differentially expressed between LMS and UPS and that they were not separable using unsupervised hierarchical clustering 10 ; an advantage of the similarity is that common prognostic genes could be identified. Similar results have been reported for gene methylation; hierarchical cluster analysis of methylation status clearly demarcated sarcomas with specific translocations (myxoid liposarcoma and synovial sarcoma) but the high grade tumors studied here (MFS, LMS and UPS) were heterogeneously dispersed through two different branches of the dendrogram 11. Interestingly, whereas a classifier could accurately identify LMS, the tumors MFS and UPS (along with pleomorphic liposarcoma) could not be clearly distinguished. MALDI MSI does not have the depth of coverage (number of proteins/genes) as gene expression profiling utilizing array or NGS technology, nor can it detect methylation states, and the differential diagnosis biomarkers referred to above were not detected in our experiments. Among the twenty protein ions found here to be potential biomarkers for the differential diagnosis (diagnostic biomarkers) of high grade sarcomas, four could be assigned using databases of protein ions commonly detected by MALDI-MSI 64,65 : thioredoxin, macrophage migration inhibitory factor (MIF), galectin-1 and acyl- CoA-binding protein. All four have previously been associated to cancer (see Tab.3.2). For example, using gene expression data Takahashi et al. reported MIF in combination with stearoyl-coa desaturase 1, to be a marker for the differential diagnosis of soft tissue sarcoma (STS) subtypes (UPS and MFS) 134. In a follow-up study it was reported that this gene expression signature can be used as a 65

73 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging prognostic biomarker for STS 127. It is also intriguing that the diagnostic protein identified by Takahashi et al. (stearoyl-coa desaturase 1) and our study (acyl-coabinding protein) are both involved in the processing of fatty acids. The diagnostic biomarkers reported were obtained from patient tissues with typical histologies for the sarcomas, as confident diagnosis is essential for the biomarker discovery pipeline. As a result they are directly applicable only to the cases that can be differentiated on the basis of histology; more work is necessary to establish how they translate to examples with a difficult differential diagnosis. In our prognostic biomarker analysis, we combined tumor subtypes to assess if a common prognostic biomarker could be detected. Among the nine protein ions found to be potential prognostic biomarkers of high grade sarcomas, associated to overall survival, four could be assigned using databases of protein ions commonly detected by MALDI-MSI 64,65 : proteasome activator complex subunit 1(PSME1), two Histone H4 variants and hemoglobin subunit beta. All four have previously been associated to cancer (see Tab.3.3). Using MALDI-MSI PSME1 has been reported as a marker of stromal activation in breast cancer 131, and as a tumor marker in human esophageal squamous cell carcinoma 135 and ovarian cancer 39,41. PSME1 has also been reported as a tumor target for prostate cancer: it was shown that intravenously injected anti-psme1 antibodies provide the means to deliver therapeutic agents into primary and metastatic prostate carcinoma 136. Soft-tissue sarcomas are rare and collecting fresh frozen tissues is not standard clinical practice in all institutes, accordingly the number of available patient samples is often limited. The work described here used four distinct filters to guard against false positive identifications: i) Screening for masses that are correlated with the clinical data using significance analysis of microarrays, which controls the false positive identification rate through a 5% false positive rate threshold (q-value <= 0.05); ii) Masses that passed the SAM filter tested were then subject to Kaplan- Meier analysis for significance (log rank test: p <= 0.05); iii) A result was only considered reliable if the total number of samples in the subgroups were greater than 10. In order to increase the number of samples for the test, we also grouped different tumor types, for instance all non-osteosarcoma patients. iv) A Cox proportional hazards regression model was used to ensure biomarkers were independently correlated with the clinical endpoint with regard to other clinical factors such as neoadjuvant therapy, tumor type, etc. 66

74 Chapter 3 Diagnostic and Prognostic Proteins We also demonstrated that focusing the prognostic biomarker analysis to specific differentiation grades was beneficial: an additional two biomarkers were found in this manner that were not statistically significant if the average mass spectrum of all differentiation grades were used instead, and those that were detected previously exhibited a greater statistical significance when the analysis was limited to specific differentiation grades. The additional two prognostic biomarkers are protein ions detected at m/z 6,281 and 6,653 (the latter assigned as cytochrome C oxidase subunit two 132 ), which were statistically associated with overall survival for high grade LMS when the analysis was focused on the undifferentiated regions. Protein signals that exhibited greater statistical significance if the analysis focused on a specific differentiation grade include m/z 15,877, assigned as Hemoglobin subunit beta 128 and indicative of poor overall survival in MFS patients; if the analysis was focused on the well differentiated regions the resulting p value was four-fold better than that obtained if all differentiation grades were used. A detailed overview of the benefits of increased histological specification is reported in Supplementary Tab.3.7; the consistent increases in statistical power reflects the reduced variation in molecular signatures (that arises from the variable cellular makeup, with respect to cellular differentiation, in different patient samples). Prognostic biomarkers not only enable the stratification of patients, they may also provide important insights into tumor progression that may ultimately lead to novel treatment strategies. For example Takahashi et al. provide a hypothetical regulation model for metabolic and signaling control from the prognostic biomarkers found in soft tissue sarcomas 127. In another example, it was demonstrated that a biomarker that could predict patient response to neoadjuvant chemotherapy in oesophageal adenocarcinoma was linked to mitochondrial defects 52 ; interestingly these same biomarkers also differentiated poor survival patients. Our next question was whether the histologically specified regions had specific subpopulations that drive tumor progression and determine the disease outcome of the patients 77. Using the protocol reported by Balluff et al. 79, in which the MSI data is first analyzed to identify subpopulations characterized by distinct biomolecular profiles and which are then compared with the patient s clinical data, we found tumor subpopulations that were significantly associated with patient overall survival (Fig.3.5). In this study though there was no candidate biomarker for metastasis-free survival 67

75 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging independent of clinically relevant co-factors (such as neoadjuvant therapy), one protein ion detected at m/z 8,093, whether the analysis focused on all tumor areas or focused on a specific differentiation grade (Fig.3.3f and Supplementary Tab.3.7), displayed a trend associating to metastasis-free survival in UPS patients (with p value of 0.03 in Kaplan-Meier analysis while p value of 0.06 in multivariate analysis). Intratumor heterogeneity is known to be an important factor that affects tumor progression and the clinical management of patients. High grade sarcomas are not only highly heterogeneous at the histological level (Fig.3.1), but we now also show heterogeneity on the basis of the molecular signatures detected by MSI, even within histologically homogeneous tumor areas (Fig.3.5). The above results demonstrate that the power of the diagnostic and prognostic biomarkers may be increased by combining the assay with histological specification. Importantly, the identification of these prognostically important tumor subpopulations will enable their analysis using in-depth omics technologies to further understand their progression and pathways that affect patient outcome. In summary, our results confirm the morphological and molecular intratumor heterogeneity in high grade sarcomas, and hence the need for a spatially-resolved read-out of the molecular information of the tumor. In our study we identified a panel of diagnostic and prognostic protein markers that could either distinguish between the different high grade sarcomas, or were associated with patient survival. 3.5 Acknowledgements The authors would like to acknowledge financial support from COMMIT, Cyttron II and the ZonMW Zenith project Imaging Mass Spectrometry-Based Molecular Histology: Differentiation and Characterization of Clinically Challenging Soft Tissue Sarcomas (No ). BB is funded by the Marie Curie Action of the European Union (SITH FP7-PEOPLE-2012-IEF No ). The authors also thank University Hospital Leuven for contributing an MFS patient sample. 68

76 Chapter 3 Diagnostic and Prognostic Proteins Supplementary information Supplementary information Supplementary Fig.3.1 Global basepeak spectrum of all samples. Supplementary Fig.3.2 Differential immunohistochemistry expression of proteasome activator complex subunit 1 (PSME1). 69

77 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Supplementary Fig.3.3 Two approaches for Intratumor heterogeneity investigation. Step Time 4% formalin 60sec tap water 3sec deionized water 3sec hematoxylin 3min deionized water 3sec ammonia water 20sec tap water 3sec deionized water 3sec eosin 15sec 100% ethanol 3sec 100% ethanol 3sec 100% ethanol 3sec xylene 3sec xylene 3sec xylene 3sec Supplementary Tab.3.1 H&E staining protocol 70

78 Chapter 3 Diagnostic and Prognostic Proteins Supplementary information Parameter Setting Resolution 800 Baseline Subtraction Top Hat Baseline 10% Minimal Baseline Width Savitzky Golay Spectra Smoothing 2 m/z Width Preparation 5 Cycles 2000ppm Maximal Peak Shift Recalibration 10% Match to Calibration Peaks Exclude not Recalibratable Spectra Peak Preparation Peak picking 5 Signal to Noise Threshold (on average spectrum) Supplementary Tab.3.2 ClinProTools settings. 71

79 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Phase Parameter Value Peak picking on sample spectra for alignment Spectrum to use Mean TIC-normalized Kaiser smoothing window [data points] 40 Resampling rate [Da] 1 M/z block [Da] 700 LIMPIC baseline subtraction uses m/z block Minimum signal-to-noise 4 Minimum half peak width 4 Max. number of reference peaks 10 Peak clustering tolerance [ppm] 2000 Minimum peak detection rate 85% Alignment Width of pulses (msalign function) 5 Resampling rate [Da] 1 Peak picking on global mass spectrum Spectrum to use Basepeak spectrum Kaiser smoothing window [data points] 40 Resampling rate [Da] 1 M/z block [Da] 700 TopHat baseline subtraction width [data points] 1000 Minimum signal-to-noise 2 Minimum half peak width 8 Peak width estimation [ppm] 1800 Read-out of data and processing Use intensity or area of peaks Area Spectrum normalization Total ion count (TIC) Remove extreme mass spectra according to TIC 1% highest and 1% lowest Offset for spatial arrangement of samples [px] 10 Supplementary Tab.3.3.Parameters for spectra processing in Matlab 72

80 Chapter 3 Diagnostic and Prognostic Proteins Supplementary information Antigen Proteasome activator complex subunit 1 (PMSE1) Macrophage migration inhibitory factor (MIF) Galectin-1 (Gal-1) Compan y ABCAM Sigma Novocas tra product code ab HPA NCL-GAL1 Antigen retrieval citrate ph 6.0 citrate ph 6.0 citrate ph min with 5% non-fat dry milk in PBS/1%BSA 30 min with 5% non-fat dry milk in PBS/1%BSA incubatio n ON at 4 ON at 4 no ON at 4 preincubation 30 min with Acyl-CoAbinding ABCAM ab citrate ph 5% non-fat 6.0 dry milk in protein (ACBP) PBS/1%BSA Supplementary Tab.3.4. Pre-optimized protocol of immunostaining. ON at 4 dilution 1:1500 in PBS/1% BSA 1:4000 in PBS/1% BSA 1:800 in PBS/1% BSA 1:5000 in PBS/1% BSA/5%nonfat dry milk. 73

81 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging m/z observed (median) Protein assignme nt Highest mean intensit y in Lowest mean intensit y in p value_r un 1 p value_run 2 p value_r un 3 p value_r un 4 p value_r un 5 Signif icanc e perce ntage 6,226 no LMS MFS N.D % 9,522 no LMS MFS % 10,374 no LMS MFS % 10,395 no LMS MFS % 11,051 no LMS MFS % 11,138 no LMS MFS % 11,162 11,608 12,350 14,633 Acyl-CoAbinding protein (ACBP) Thioredo xin (Trx) Macroph age migration inhibitory factor (MIF) Galectin- 1 (Gal-1) LMS LMS MFS MFS % 100% LMS MFS N.D % LMS MFS % 17,093 no LMS MFS % 17,419 no LMS MFS N.D. 80% 17,632 no LMS MFS % 17,832 no LMS MFS % 19,929 no LMS MFS N.D % 20,729 no LMS MFS ,464 no LMS MFS % 22,472 no LMS MFS % 22,677 no LMS MFS ,752 no LMS MFS % Supplementary Tab protein peaks (p 0.05) from diagnostic biomarker analysis. 100% 100% 74

82 Chapter 3 Diagnostic and Prognostic Proteins Supplementary information m/z observe d Protein assignment Subdataset 4,680 no UPS 9,753 Proteasome activator complex subunit 1 Histone H4 (a) Histone H4 (b) Non-OS Analysis Overall survival Overall survival Patient number p value of Kaplan- Meier analysis p value of multivariate analysis E ,314 LMS Overall survival ,355 LMS Overall survival ,827 no Non-OS Overall survival ,856 no Non- Overall 0.044/ / /11 OS/LMS survival ,909 no LMS Overall survival ,877 Hemoglobin Overall no co-factor MFS subunit beta survival variance 16,084 no MFS Overall no co-factor survival variance Supplementary Tab.3.6. Nine protein peaks (p <= 0.05) from prognostic biomarker analysis 75

83 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging m/z observed 8,093 no 15,877 Assignment Analysis Subdataset Hemoglobin subunit beta 16,084 no 6,281 no 6,653 Cytochrome c oxidase subunit 2 Metastasisfree survival Overall survival Overall survival Overall survival Overall survival Patient number p value of Kaplan- Meier analysis p value of Multivariate Analysis UPS UPS_undiff MFS MFS MFSwelldiff. MFSwelldiff no co-factor variance no co-factor variance no co-factor variance no co-factor variance LMS_undiff LMS_undiff Supplementary Tab.3.7. Prognostic biomarkers detected from tumor areas with specific differentiation grades. Undiff.: undifferentiated tumor area; welldiff: well differentiated tumor area 76

84 Chapter 4 High nuclear expression of proteasome activator complex subunit 1 predicts poor survival in soft tissue leiomyosarcomas Lou S, Cleven AH, Balluff B, de Graaff M, Kostine M, Briaire-de Bruijn I, McDonnell LA, Bovée JV. Clin Sarcoma Res. 2016; 6: 17. Published online 2016 Oct 1. doi: /s z

85 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Abstract Background: Previous studies on high grade sarcomas using mass spectrometry imaging showed proteasome activator complex subunit 1 (PSME1) to be associated with poor survival in soft tissue sarcoma patients. PSME1 is involved in immunoproteasome assembly for generating tumor antigens presented by MHC class I molecules. In this study, we aimed to validate PSME1 as a prognostic biomarker in an independent and larger series of soft tissue sarcomas by immunohistochemistry. Methods: Tissue microarrays containing leiomyosarcomas (n=34), myxofibrosarcomas (n=14), undifferentiated pleomorphic sarcomas (n=15), undifferentiated spindle cell sarcomas (n=4), pleomorphic liposarcomas (n=4), pleomorphic rhabdomyosarcomas (n=2), and uterine leiomyomas (n=7) were analyzed for protein expression of PSME1 using immunohistochemistry. Survival times were compared between high and low expression groups using Kaplan- Meier analysis. Cox regression models as multivariate analysis were performed to evaluate whether the associations were independent of other important clinical covariates. Results: PSME1 expression was variable among soft tissue sarcomas. In leiomyosarcomas, high expression was associated with overall poor survival (p = 0.034), decreased metastasis-free survival (p = 0.002) and lower event-free survival (p = 0.007). Using multivariate analysis, the association between PSME1 expression and metastasis-free survival was still significant (p = 0.025) and independent of the histological grade. Conclusions: High expression of PSME1 is associated with poor metastasis-free survival in soft tissue leiomyosarcoma patients, and might be used as an independent prognostic biomarker. Keywords proteasome activator complex subunit 1; prognostic biomarker; sarcoma; leiomyosarcoma; soft tissue sarcoma; immunohistochemistry Abbreviations PSME1: Proteasome activator complex subunit one; LMS: Leiomyosarcomas; MFS: Myxofibrosarcomas; UPS: Undifferentiated pleomorphic sarcomas; Uterine LM: Uterine leiomyomas; LPS: Pleomorphic liposarcomas; RMS: Pleomorphic rhabdomyosarcomas; USCS: Undifferentiated spindle cell sarcomas 78

86 Chapter 4 Immunohistochemistry Validation 4. High Nuclear Expression of Proteasome Activator Complex Subunit 1 Predicts Poor Survival in Soft Tissue Leiomyosarcomas 4.1 Background Soft tissue sarcomas are a heterogeneous group of rare malignancies often having poor outcome 1. Soft tissue sarcomas constitute less than 1% of all cancers 1 while there are more than 50 histological subtypes with sometimes overlapping histological features 3. Distinction is essential as subtypes differ in biological behaviour and sensitivity to chemotherapy, and as such an adequate histological diagnosis, is crucial for clinical decision making 6. Fifty-six % of soft tissue sarcomas present as localized disease at the time of diagnosis, and surgery is the mainstay of treatment, sometimes combined with radiotherapy or chemotherapy 8. From the molecular point of view, soft tissue sarcomas can be distinguished into two categories. The first class includes sarcomas with a simple genome, in which recurrent translocations, amplifications or specific mutations can be found. The second class includes sarcomas with a complex genome, characterized by a multitude of chromosomal alterations and genomic instability, often reflected by pleomorphic histological features 6. This group includes high grade leiomyosarcoma, myxofibrosarcoma, undifferentiated pleomorphic sarcoma, undifferentiated spindle cell sarcoma, pleomorphic liposarcoma, and pleomorphic rhabdomyosarcoma. Leiomyosarcomas constitute 5-10% of all soft tissue sarcomas, displaying smoothmuscle differentiation 1. Studies showed for leiomyosarcoma that the metastasisfree 5-year survival rate is about 60% 7. Histological grade is the most important prognostic factor for most soft tissue sarcomas. By using FNCLCC grading system, which is the most widely used 3-grade system, soft tissue sarcomas are divided into low, intermediate and high grade based on the sum score of three histologic parameters including tumor differentiation, mitotic count and tumor necrosis. About 65% of leiomyosarcomas are reported to have high-grade areas 137. High grade leiomyosarcomas often have poor patient outcome 8. Until now, the genetics and pathology of leiomyosarcomas are not completely understood and as they have a complex genome, no molecular diagnostic tests or specific therapeutic targets are available. Hence, there is a strong need for new molecular markers that can aid in the stratification of leiomyosarcomas patients with respect to their 79

87 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging disease outcome. In a previous study, we used imaging mass spectrometry to compare these soft tissue sarcomas with a complex genome. A panel of protein signatures that could distinguish between different subtypes, or were associated to patient survival were discovered 102. Among them, proteasome activator complex subunit 1 (PSME1) was found indicative of poor survival in soft tissue sarcomas. PSME1 (also known as REGalpha and PA28A), is a multicatalytic proteinase complex, implicated in immunoproteasome assembly and required for efficient antigen processing 138. Intriguingly, PSME1 was also found to associate with diagnosis or prognosis in other tumor types, e.g. prostate cancer 136, breast cancer 131 and ovarian cancer 39,41. In this study, we used tissue microarrays of soft tissue sarcomas with complex genomes, to evaluate whether PSME1 expression can predict clinical outcome in soft tissue sarcomas, especially leiomyosarcomas. 4.2 Methods Tissue Microarrays Tissue microarrays were previously constructed from paraffin embedded formalin fixed tissues using a semi-automated TMA apparatus (TMA Master; 3D Histech, Budapest, Hungary) 139. Clinicopathological details were described previously 140. In brief, analysed samples include 34 leiomyosarcomas, 14 myxofibrosarcomas, 15 undifferentiated pleomorphic sarcomas, 4 undifferentiated spindle cell sarcomas, 4 pleomorphic liposarcomas, 2 pleomorphic rhabdomyosarcomas, and 7 uterine leiomyomas. Clinicopathological data for the leiomyosarcomas, as described previously 140, are summarized in Supplementary Tab.4.1. All tumor samples are present at least in triplicates with a diameter of 1.5 mm (a surface area of around mm 2 ). Cores from colon, liver, placenta, prostate, skin, and tonsil were included for control and orientation purposes. Four micrometre thick sections were transferred by using a tape-transfer system to coated glass slides for analysis. The histological diagnosis of all samples was confirmed by reviewing the hematoxylin and eosin stained slides by expert pathologist (J. V. M. G. B.). Malignant tumors were graded according to the FNCLCC (La Fédération Nationale des Centres de Lutte Contre le Cancer) grading system 1. All samples were handled according to the Dutch code of proper secondary use of human material as accorded by the Dutch society of pathology (Federa). The samples were handled in a coded manner. All study methods were approved by the LUMC ethical board 80

88 Chapter 4 Immunohistochemistry Validation (B16.025) PSME1 immunohistochemistry Four micrometre thick sections were dried overnight at 37 C. Immunohistochemistry was performed using anti-psme1 antibody (clone [EPR10968(B)], abcam, Cambridge, UK) according to protocols described previously 141. Briefly, slides underwent deparaffinization, blocking of endogenous peroxidase, antigen retrieval (10 min microwave in citrate, ph 6.0), pre-incubation, and addition of the primary antibody in a dilution of 1:1500 overnight. Next, slides were incubated with Poly-HRP-GAM/R/R (Immunologic BV, Duiven, The Netherlands (DPVO110HRP)), visualized with DAB+Substrate Chromogen System (DAKO, Heverlee, Belgium) and counterstained with hematoxylin. Colon tissue was used as a positive control. As a negative control slides were incubated with PBS/1 % BSA instead of the primary antibody Scoring of immunohistochemistry Slides were scored independently by two observers (J.V.M.G.B and A.H.G.C) as described previously 142. In brief, staining intensity (0, absent; 1, weak; 2, moderate; 3, strong) and percentage of positive tumor cells (0, 0%; 1, 1 24%; 2, 25 49%; 3, 50 74%; 4, %) were assessed. Afterwards, scores of staining intensity and percentage of positive tumor cells were added to obtain the sum score; for later statistical analysis, the average sum score was calculated over all cores belonging to the same tumor. Proteasomes are present both in the nucleus as well as in the cytoplasm of eukaryotic cells, although their relative abundance within these compartments can be highly variable 138, We therefore evaluated cytoplasmic and nuclear staining separately. Cores in which tissue was lost or with not enough tumor area were excluded from the analysis. Cores with differences on sum score from two observers more than 2 were re-evaluated to reach consensus Statistical analysis Only primary tumour samples were used in statistical analysis. First, the distribution of sum score data was evaluated by Shapiro-Wilk normality test. As this test showed that the score data was not normally distributed, nonparametric Spearman correlation coefficient was used as a measure of the statistical 81

89 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging dependence between the histological grades and PSME1 expression. Further statistical two-group comparisons between controls (uterine leiomyoma) and the different histological grades of soft tissue sarcomas were calculated by Dunn s multiple-comparison test. Spearman correlation was performed in R environment (R Foundation for statistical Computing, Vienna, Austria), scatter plots and Dunn s test results were generated in GraphPad Prism version 6.00 for Windows (GraphPad Software, La Jolla, California, USA, All two-sided p values equal or lower than 0.05 were considered statistically significant. For survival analysis, patients were dichotomized into two groups. We dichotomized leiomyosarcoma patients into high and low expression groups according to the sum scores of immunohistochemistry, for which we chose the cut-off at the 3rd quartile Experience shows that molecular subgroups are usually found in 10% to 25% of the patients (e.g. HER2 overexpression 114, KRAS mutation 115 ). Differences in overall survival, metastasis-free survival and eventfree survival between these groups were investigated using Kaplan-Meier curves and the log-rank test. Independent variables predicting survival were evaluated in a multivariable model using Cox Regression analyses. Survival analysis was performed in R environment (R Foundation for statistical computing, Vienna, Austria) using Survival package and all two-sided p values lower or equal than 0.05 were considered statistically significant. 4.3 Results Variable nuclear and cytoplasmic expression of PSME1 in soft tissue sarcomas In soft tissue sarcomas, PSME1 protein expression was found in the majority of the cases, both in the nucleus as well as in the cytoplasm. In contrast, expression in benign leiomyoma was low or absent (Fig.4.1a-b). Representative images of immunohistochemistry are shown in Fig Increased expression of PSME1 with increasing histological grade in leiomyosarcomas The leiomyosarcoma subgroup was large enough to analyse a possible correlation with histological grade. Indeed, while expression was low to absent in uterine leiomyoma, expression gradually increased with increasing histological grade in 82

90 Chapter 4 Immunohistochemistry Validation both nucleus (p overall = ) and cytoplasm (p overall = ) in leiomyosarcomas (Fig. 4.1c-d). Further statistical two groups comparisons between control and any histological grade by Dunn s multiple comparisons test showed that both nuclear and cytoplasmic staining significantly differed in uterine leiomyomas versus leiomyosarcomas grade 2 (p 0.05) and uterine leiomyomas versus leiomyosarcomas grade 3 (p 0.01). Fig.4.1 Summary of PSME1 immunohistochemistry results. Variable expression of PSME1 both in the cytoplasm (a) as well as in the nucleus (b) in soft tissue sarcomas, while expression in uterine leiomyoma (LM; control) is low. LMS: leiomyosarcomas, LPS: pleomorphic liposarcomas, MFS: myxofibrosarcomas, RMS: pleomorphic rhabdomyosarcomas, UPS: undifferentiated pleomorphic sarcomas, and USCS: undifferentiated spindle cell sarcomas. In leiomyosarcomas, both cytoplasmic (c) and nuclear (d) expression increased with increasing histological grade. (p = and p = ). In addition, both cytoplasmic and nuclear expression of PSME1 was significantly higher in intermediate and high grade leiomyosarcomas as compared to uterine leiomyomas (p 0.05/ p 0.01). All score data for each group were presented in 83

91 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging mean ± SD. Fig.4.2. Representative images of immunohistochemistry of PSME1. (a) and (b) are two leiomyosarcoma (LMS) samples with high expression of PSME1. (c) is a uterine leiomyoma (LM) control sample with low expression of PSME1. Images in red squares are the overviews of expression the tissue microarray cores for respective samples High nuclear expression of PSME1 predicts poor outcome in leiomyosarcoma patients To investigate a possible correlation of PSME1 expression with clinical outcome, leiomyosarcoma patients were dichotomized into high and low PSME1 expression groups according to the sum scores of immunohistochemistry. High PSME1 expression was associated with poor overall survival (p = 0.034), decreased metastasis-free survival (p = 0.002) and lower event-free survival (p = 0.007) (Fig.4.3). Fig.4.3. Kaplan-Meier survival plots of PSME1. Kaplan-Meier plots comparing the different survival data of leiomyosarcoma patients with respect to a high and low nuclear expression of PSME1 (cut-off: 3rd quartile). High nuclear expression of PSME1 in 84

92 Chapter 4 Immunohistochemistry Validation leiomyosarcoma was significantly associated with decreased overall survival, metastasisfree survival and event-free survival (log-rank test; p 0.05) High nuclear expression of PSME1 as an independent prognostic factor in leiomyosarcoma patients Using multivariable Cox Regression analyses including clinically relevant co-factors such as histological grade, age and gender, we showed that high nuclear expression of PSME1 was independently associated with metastasis-free survival (p = 0.03) (Tab.4.1). The independent predictive power of nuclear PSME1 expression for overall and event-free survival was at the border of significance (p = 0.07) (Tab.4.1). Clinical 95% Confidence Variable Hazards Ratio association Interval p-value Metastasis-free survival PSME1 high nuclear expression Histological grade Age Gender (M) Event-free survival PSME1 high nuclear expression Histological grade Age Gender (M) Overall survival PSME1 high nuclear expression Histological grade Age Gender (M) Tab.4.1. Results of multivariable analysis of factors influencing survival 85

93 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging 4.4 Discussion Using imaging mass spectrometry we previously identified PSME1 as a prognostic biomarker indicating poor survival in soft tissue sarcoma patients 102. Imaging mass spectrometry is a sensitive discovery tool (zepto-molar sensitivity 147 ) enabling the detection of hundreds of molecules directly from tissue 25,85. To further explore the prognostic value of PSME1 we analysed PSME1 expression in a larger, independent set of soft tissue sarcomas using immunohistochemistry on tissue microarrays. PSME1 (or PA28A) encodes a subunit of the proteasome system, which is a major source for generation of tumor antigens presented by MHC class I molecules 148,149. Escape of immune response is one of the hallmarks of cancer 57. In addition, elevated proteasome activity in tumor cells has been described to influence transcription factors involved in cell survival or apoptosis 150,151. Novel strategies using the proteasome have been proposed for cancer treatment for example by alternating the NAD+/NADH ratio to change kinetics of proteasomal degradation 151 or inhibiting proteasome to induce apoptosis PSME1 is expressed in many different cell types, especially antigen presenting cells, and its expression can be controlled by interferon gamma. Both chemotherapy and TNF-alpha may induce a local inflammatory reaction within the tumor microenvironment and therefore may influence expression of PSME1. It is of interest that all sarcoma subtypes included in our study expressed PSME1 to a variable extent, while neoadjuvant chemotherapy or treatment with interferon gamma is not standard practice in our hospital. As far as clinical data were available, only four patients received preoperative chemotherapy or TNF-alpha, and expression levels were not significantly different. In the control group, consisting of uterine leiomyomas, expression was low to absent, both in the nucleus as well as in the cytoplasm. High PSME1 expression was also described in other tumors. For example, increased PSME1 expression was also found in primary and metastatic human prostate cancer and was suggested as a potential target for therapeutic intervention 136. PSME1 was previously also detected using imaging mass spectrometry in other tumors: Dekker et al. detected PSME1 as a marker of stromal activation in breast cancer 131. Previous studies also showed that PSME1 could be a molecular signature to discriminate between benign and malignant ovarian tumors 39,40, and an early diagnosis and tumor-relapse biomarker 41. Zhang et al. detected PSME1 as a tumor marker in human oesophageal squamous cell carcinoma 135. The proteasome can be present in the cytoplasm as well as in the 86

94 Chapter 4 Immunohistochemistry Validation nucleus of all eukaryotic cells, although their distribution and function can be variable 143. We here show that in soft tissue sarcomas with a complex genome, PSME1 is expressed both in the cytoplasm and in the nucleus. Proteasomedependent protein degradation is important in the cytoplasm for MHC class 1 antigen presentation 138. In the nucleus, PSME1 plays an important role in maintaining the nuclear function including gene expression and cell proliferation 145,156. To further evaluate its clinical relevance, we analysed the largest subgroup, comprising 34 leiomyosarcomas of different histological grade, in more detail. Both nuclear as well as cytoplasmic expression of PSME1 significantly increased with increasing histological grade. Moreover, high nuclear expression of PSME1 was significantly associated to poor outcome (overall survival, metastasis-free survival and event-free survival) in leiomyosarcoma patients, although the patient cohort is rather small (n=34). In multivariate analysis only the association with decreased metastasis-free survival was independent of histological grade, while an independent association to poor overall survival and decreased event-free survival was at the border of significance. Although PMSE1 expression is a promising biomarker, our results need to be validated in an independent cohort of leiomyosarcomas. In summary, we found elevated expression of the proteasome subunit PSME1 in leiomyosarcomas compared to control tissues, and an association of the expression with increasing histological grade in leiomyosarcoma. Moreover, high nuclear PSME1 expression was found to be an independent predictor of metastasis-free survival in leiomyosarcoma patients. Our results suggest that the expression of proteasome subunits such as PSME1 could be taken into account for leiomyosarcoma patients when considering immunotherapeutic strategies in these tumors Conclusions We show variable expression of PSME1 in different soft tissue sarcoma subtypes with complex genomes. Our results showed that high nuclear expression of proteasome activator complex subunit 1 is an independent poor prognostic factor in leiomyosarcomas, which suggests that the proteasome could be exploited as a possible novel target for the treatment of leiomyosarcomas. 87

95 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging 4.6 Acknowledgements The authors would like to thank B.E.W.M. van den Akker for technical assistance. 88

96 Chapter 4 Immunohistochemistry Validation Supplementary information Supplementary information Feature Value No. patients 34 Age Median (years) 63 Range (years) Gender Male : Female 13: 21 Histological grade Grade 1: Grade 2: Grade 3 6: 9: 19 Length of follow-up Median metastasis-free survival (months) 56.9% at max.follow-up length Median event-free survival (months) 87.2 Median overall survival (months) 80.2 Supplementary Tab.4.1. Clinicopathologic characterization of leiomyosarcoma patients

97 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging 90

98 Chapter 5 Prognostic metabolite biomarkers for soft tissue sarcomas discovered by mass spectrometry imaging Lou S, Balluff B, Cleven AH, Bovée JV, McDonnell LA. J. Am. Soc. Mass Spectrom. 2017; 28(2): Published online 2016 Nov 21. doi: /s

99 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Abstract Metabolites can be an important read out of disease. The identification and validation of biomarkers in the cancer metabolome that can stratify high-risk patients is one of the main current research aspects. Mass spectrometry has become the technique of choice for metabolomics studies, and mass spectrometry imaging (MSI) enables their visualization in patient tissues. In this study we used MSI to identify prognostic metabolite biomarkers in high grade sarcomas; 33 high grade sarcoma patients, comprising osteosarcoma, leiomyosarcoma, myxofibrosarcoma and undifferentiated pleomorphic sarcoma were analyzed. Metabolite MSI data was obtained from sections of fresh frozen tissue specimens with matrix-assisted laser/desorption ionization (MALDI) MSI in negative polarity using 9-aminoarcridine as matrix. Subsequent annotation of tumor regions by expert pathologists resulted in tumor-specific metabolite signatures, which were then tested for association with patient survival. Metabolite signals with significant clinical value were further validated and identified by high mass resolution Fourier transform ion cyclotron resonance (FTICR) MSI. Three metabolite signals were found to correlate with overall survival (m/z and m/z ) and metastasis-free survival (m/z ). FTICR- MSI identified m/z as inositol cyclic phosphate and m/z as carnitine. Keywords Metabolites; prognosis; biomarker discovery; MALDI-MSI; high grade sarcoma; leiomyosarcoma; myxofibrosarcoma; osteosarcoma; undifferentiated pleomorphic sarcoma; soft tissue sarcoma Abbreviations MSI: Mass spectrometry imaging; MALDI: Matrix-assisted laser/desorption ionization; FTICR: Fourier transform ion cyclotron resonance; DESI: Desorption electrospray ionization; MFS: Myxofibrosarcoma; LMS: Leiomyosarcoma; UPS: Undifferentiated pleomorphic sarcoma; OS: Osteosarcoma; H&E: Hematoxylin and eosin; ROI: Regions-of-interest; TIC: Total-ion-count; SAM: Significance analysis of microarrays 92

100 Chapter 5 Prognostic Metabolites 5. Prognostic Metabolite Biomarkers for Soft Tissue Sarcomas Discovered by Mass Spectrometry Imaging 5.1 Introduction Soft tissue sarcomas are a rare group of tumors, comprising less than 1% of all malignant tumors 158 with more than fifty histological subtypes 159. About 10% of patients with soft tissue sarcoma have detectable metastases (most common in the lungs) at diagnosis of the primary tumor, and at least one third of patients die from tumor-related disease (most of them from lung metastases) 158. Additionally, different tumor subpopulations can have different clinical-pathological behavior 4,73,102. The combination of rarity, many subtypes with overlapping and heterogeneous histologies, and molecular intratumor heterogeneity have made the correct diagnosis and treatment of high-grade sarcomas challenging. Different therapeutic strategies have been developed but the choice is dependent on subtype because they confer different chemosensitivity 100. Accurate histological classification of sarcomas can be far from straightforward: full concordance between primary diagnosis and second opinion was observed in less than 60% 160. Additionally, prognostic factors associated with local failure and overall survival are also dependent on tumor location and size 158. There is an urgent need for more objective molecular predictors to improve prognosis and treatment. A cell s metabolome has been defined as the unique chemical fingerprints left behind by a cell s specific molecular process 161 and altered metabolism is now one of the hallmarks of cancer 57. Cancer cells can exhibit significant alterations in metabolic pathways such as glycolysis, respiration, the tricarboxylic acid cycle, oxidative phosphorylation, lipid metabolism, and amino acid metabolism 162. Tumors show various metabolic aberrations but perhaps the most central to tumor proliferation is the Warburg effect 59. Cancer cells reprogram their energy metabolism by limiting energy metabolism to glycolysis, producing excessive levels of lactate, even under normoxic conditions ( aerobic glycolysis ). The high rate of glucose uptake and glycolysis in tumors is the basis for the widely used tumor imaging technique [18F] deoxyglucose-positron emission tomography (18FDG PET) The molecular basis for increased glycolysis and altered metabolism is currently an active area of onco-metabolomic research, and spans the identification and validation of biomarkers in the cancer metabolome that can stratify high-risk 93

101 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging patients and/or distinguish between benign and advanced metastatic forms of the disease to the functional consequences of the altered metabolic states 166,167. Previous studies on sarcoma-metabolomics has mostly concerned cell lines because of the much greater ease of freezing their metabolome status; the metabolome reacts quickly to changes in oxygen, blood supply etc. Nevertheless, it is important to assess metabolites within the histopathological context of patient tumor samples, especially so for histologically heterogeneous tumors such as sarcomas. Mass spectrometry imaging (MSI) has had a rapid and substantial impact on clinical research 171 because of its ability to directly visualize a tissue s molecular content without any labeling or a-priori knowledge 28,85,86. MSI enables, through a seamless integration of histology with the MSI data, cell-type specific molecular signatures to be obtained from the real histopathological context of patient tissues 25,97. Both matrix assisted laser desorption/ionization (MALDI) and desorption electrospray ionization (DESI) have been used to acquire metabolite MSI datasets of cancer tissues. For example MALDI MSI may be used to visualize metabolites and metabolic pathways in fresh frozen tumor tissues 172 and identify biomarkers including in formalin fixed paraffin embedded tissues and tissue microarrays 173 ; DESI has been used to identify metabolite biomarkers for discriminating between different brain and breast tumors 174,175. Combining tumorspecific metabolite signatures with patient follow-up data has revealed prognostic biomarkers 173. Though less established for clinical research than protein MSI these studies indicate that the metabolite signatures obtained via MSI can also be used to differentiate between different tumor entities and between patient groups. MSI has previously been applied in sarcomas. Caldwell et al. used MALDI MSI to investigate protein changes in the tumor microenvironment of malignant fibrous histiocytoma 176. MALDI MSI of proteins has revealed molecular intratumor heterogeneity in myxofibrosarcoma 73, and it has been used to determine proteins and lipids that are tumor-type and tumor-grade specific for myxoid liposarcoma and myxofibrosarcoma 4. However an assessment of the prognostic value of metabolites for soft tissue sarcomas has not been reported to date. In this study we investigated if the metabolite and lipid signatures detected by MALDI MSI using 9 aminoacridine in high-grade sarcomas including high-grade leiomyosarcoma, myxofibrosarcoma, undifferentiated pleomorphic sarcoma and osteosarcoma, can be associated with patient prognosis. 94

102 Chapter 5 Prognostic Metabolites 5.2 Experimental Tissue specimens and sample cohorts Fresh frozen tumor samples of high-grade myxofibrosarcoma (MFS), leiomyosarcoma (LMS), undifferentiated pleomorphic sarcoma (UPS) and osteosarcoma (OS) were collected and handled as described previously 102. Briefly, tissue samples were obtained from the archive of the Department of Pathology of Leiden University Medical Center (LUMC), the Netherlands. All tumor samples were acquired during routine patient care and were handled in a coded manner according to the ethical guidelines described in Code for Proper Secondary Use of Human Tissue in the Netherlands of the Dutch Federation of Medical Scientific Societies. Slides were reevaluated histologically and classified according to the 2013 World Health Organization criteria. MFS and LMS cases were histologically graded according to the La Fédération Nationale des Centres de Lutte Contre le Cancer. The final cohort comprised 8 LMS, 10 MFS, 8 UPS and 7 OS patients (n total =33). Tab. 5.1 shows the clinic-pathological data of the patient series. No. of patients High grade LMS High grade MFS High grade UPS High grade OS Gender Male vs. Female 4 vs. 4 3 vs. 7 6 vs. 2 4 vs. 3 Age Median (y) Therapy Length of follow-up Neoadjuvant treatment Adjuvant treatment Median overall survival (mo) Median metastasis-free survival (mo) 1 / 8 0 / 10 1 / 8 5/ 7 7 / 8 6 / 10 5 / 8 6 / % survival probability at max. follow-up time (103.2 mo) 73% metastasis-free probability at max. follow-up time (52.0 mo) Tab.5.1. Clinic-pathological characteristics of the patient series

103 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Tissue preparation A semi-supervised block randomization was used to distribute the patient tissue sections between and within slides in order to minimize any potential sources of bias during MSI data acquisition (pseudo-code available in Carreira et al 177 ). Twelve micron-thick sections were cut at -20 C in a cryostat and thaw mounted onto indium-tin-oxide glass slides (Bruker Daltonik, Bremen, Germany) that had previously been coated with poly-l-lysine 178. The mounted tissue sections were stored at -80 C until use. Slides were first dried in a freeze dryer for 15 min then fiducial markers added near to the tissues using a Tipp-Ex pen. Tissue sections were sprayed with 10 mg/ml 9-aminoacridine in 70% methanol using the ImagePrep spraying device (Bruker Daltonik). The spraying method consisted of 22 cycles of 0.1 V for matrix layer thickness at 20% power and 10% modulation with 1.5 s for nebulization, 10 s of incubation, and 30 s of drying. Prior to the MSI data acquisition the matrix-coated slides were placed in the MSI slide holder and scanned with 2400 dpi resolution (Epson V200 Photo). Note the resolution of the scanned image was much higher than that of the MSI data (100 µm pixel size corresponds to 254 dpi) to enable accurate co-registration with the histological image MSI data acquisition All MSI experiments were performed using an Ultraflextreme III MALDI-ToF/ToF mass spectrometer (Bruker Daltonik) with a lateral resolution of 100 µm, a laser focus setting of large (corresponding to an on-target laser spot size of approximately 80 µm) and 500 accumulated laser shots per pixel (10 laser shots per step of a random walk within each pixel). Negatively charged ions up to m/z 1,000 were detected with a digitization rate of 4 GHz in reflectron mode. All pixel mass spectra were processed with a smoothing algorithm (SavizkyGolay algorithm, width m/z, 2 cycles) and a background subtraction (TopHat algorithm) during data acquisition using FlexAnalysis (version 3.4, Bruker Daltonik). After MSI data acquisition the slides were washed in 70% ethanol to remove the remaining matrix and then stained with hematoxylin and eosin (H&E). Highresolution digital images of the H&E stained tissues were then recorded using a Pannoramic MIDI slide scanner (3DHISTECH Ld., Budapest, Hungary) and coregistered to the MSI datasets in FlexImaging 3.0 (Bruker Daltonik). 96

104 Chapter 5 Prognostic Metabolites Histological annotation and MSI data quality control The H&E stained images were histological annotated by expert pathologists. To ensure only comparable regions of tissue were used in the subsequent statistical analysis tumor regions were classified according to their degree of differentiation (well-, moderately-, or undifferentiated). These annotated regions-of-interest (ROIs) were then used to extract the tumor-specific metabolic signatures from the MSI data. The quality of the MSI data from each ROI was then evaluated in ClinProTools 3.0 (Bruker Daltonik) by using the mass spectral preprocessing settings as reported in Supplementary Tab.5.1: any dataset in which 40% of the spectra were excluded by ClinProTools built-in quality control metrics (either being non-alignable or a null spectrum) was disqualified for the further analysis. Note: Owing to the strong presence of background peaks in the lower mass region, all ions up to m/z range 100 were excluded from the analysis MSI data processing The spectra from the annotated tumor areas of datasets that passed all quality filters embracing sample selection based on tumor cell fraction, MSI spectral quality assessment, and tissue integrity after H&E (Supplementary Tab.5.2), were loaded into MATLAB R2013a (MathWorks, Natick, MA, USA) for further processing (all settings are summarized in Supplementary Tab.5.3). There all spectra were normalized to their total-ion-count (TIC), and pixels with the 1% lowest and 1% highest TICs were discarded. Then a conservative peak picking using the LIMPIC package was performed on each sample s average spectrum upon baseline removal (TopHat) and smoothing (Kaiser) 124. Frequent peaks, which appear in over 98% of all samples, were used for recalibration of all spectra using the msalign routine (Bioinformatics toolbox) in MATLAB. After alignment, all tumor areas of the same differentiation degree were aggregated for each sample into basepeak mass spectra 113. In a next step, all basepeak spectra were averaged across all samples to obtain the global basepeak mass spectrum in which metabolite peaks were detected by the mspeaks function (Bioinformatics toolbox; MATLAB) with a minimum relative basepeak intensity threshold of 0.4%. This final project-specific peak list was then used to extract the areas-under-the-curve (AUC) values from the average spectra of each patient s tumor areas in two ways: first, to get the overall tumor profile for each patient and second, to get the overall profile 97

105 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging for each tumor differentiation degree for each patient Statistical association of metabolic signals to survival of patients The association of MSI signals of all subsets with the follow-up data was investigated similarly as in our previous study of protein biomarkers of soft tissue sarcomas 102 in R (R Foundation for statistical Computing, Vienna, Austria). Briefly, significant relationships between peak intensities and the overall/metastasis-free survival were first screened by using the Significance Analysis of Microarrays analysis in R (SAMR package). Metabolite ions with significant prognostic value (q 0.05; minimum median FDR) were then used to dichotomize patients based on the third quartile of intensity (data distribution and cutoff for one mass shown in Supplementary Fig.5.1). Differences in the overall/metastasis-free survival between the two resulting groups were compared by Kaplan-Meier curves and the log rank test using the survival package. P-values 0.05 were considered significant Metabolite identification using high mass accuracy MALDI-FTICR Additional tissue sections were prepared from patients whose MALDI-TOF-based MSI datasets contained the prognostic metabolite ions in high abundance (see above description). These tissue sections were then analyzed using an ultrahigh mass resolution mass spectrometer, namely a 9.4T MALDI Solarix XR Fourier transform ion cyclotron resonance mass spectrometer equipped with a dynamically harmonized ParaCell TM (Bruker Daltonik, Bremen, Germany). MSI was performed in negative mode using 500 laser shots per spot, laser frequency 1 khz, and 100 µm pixel size. Data was acquired from m/z 50 to m/z 1000 with a 512k data point transient and an estimated resolution of 19,000 at m/z 400 Da. Data acquisition was performed using ftmscontrol (Bruker Daltonik). Metabolite identities were assigned on the basis of accurate mass and matched isotope distributions measured with a high field FTICR MS. The metabolites measured with the FTICR were assigned to peaks measured with the MALDI-TOF system on the basis of accurate mass, isotope patterns and used matched spatial distributions as an additional constraint. 98

106 Chapter 5 Prognostic Metabolites 5.3 Results and Discussion Here we investigated whether MSI could identify metabolite biomarkers associated with survival and metastasis in high grade sarcomas. Prognostic metabolite signals were first identified based on MALDI-ToF-MSI data, then confirmed and identities assigned using high resolution, high mass accuracy MALDI-FTICR-MSI. High grade sarcomas can be histologically highly heterogeneous, with a single tumor containing areas of different histological grade, necrotic regions, and a variable inflammatory cell infiltrate. This heterogeneity can introduce high measurement variance, complicating the search for molecular biomarkers 102. Therefore, a histopathology-defined data analysis approach was utilized in which each tissue s histology was first annotated by pathologists specialized in soft tissue sarcomas. Each tissue section was first analyzed by MALDI MSI and then H&E stained. High resolution histologic images were recorded using a digital slide scanner, which were then registered to the MSI datasets using fiducial markers in FlexImaging. A virtual microdissection was then performed to isolate the mass spectra from designated tumor regions of interest. In this manner the spectra from areas with comparable histologic grade (well differentiated, moderately differentiated, undifferentiated), free of inflammatory cell infiltrate, free of necrosis, and with high tumor cell content were extracted from each patient tissue. This cell-specific data was used for the statistical analysis. The workflow is depicted in Fig.5.1. Fig.5.1. Study workflow. 1) Fresh frozen tissues with diagnosis of leiomyosarcoma, myxofibrosarcoma, undifferentiated pleomorphic sarcoma and osteosarcoma were collected and revised by pathologists in LUMC; 2) Tissue samples were cryosectioned and 99

107 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging the matrix 9-arminoarcridine applied; 3) MALDI-MSI datasets were acquired using a MALDI-TOF platform and spectra from regions of interest were extracted for analysis; 4) Datasets QC and inclusion criteria (full criteria listed in Supplementary Table 2); 5) Spectra from all the datasets were processed in MATLAB including sample alignment; 6) Masses that showed significant relationship between peak intensity and survival time were first selected using significance analysis of microarrays then were analyzed using Kaplan- Meier survival analysis, and finally were evaluated to ensure biomarkers were independently correlated with the survival time; 7) Representative poor and good survival patient samples were measured using a MALDI-FTICR platform for high accuracy; 8) Significant masses with high accuracy were assigned based on human metabolome database and using isotope patterns as an additional constraint. To control false positive identification of biomarkers we used three sequential filters: i) A Significance Analysis of Microarrays (SAM) 179 was used to screen for survival-associated peaks. SAM controls the false positive identification rate and was set here to 5%. ii) A log rank test on the dichotomized patient survival data was used to confirm the prognostic capability of the metabolite peaks. iii) Metabolite biomarker ions identified only detected in small groups of patients, less than 10, were omitted. The clinical covariates (Tab.5.1) did not exhibit any significant association to the patient survival data (age, gender, tumor type, adjuvant therapy; in all case the log-rank test resulted with P > 0.05), or the clinical covariate was significant but defined groups of too small sample size (e.g. only two neoadjuvantly treated patients) that a reliable statistical analysis could not be guaranteed. Similarly the prognostic biomarkers were not significantly associated with the clinical covariates. The analysis revealed three metabolite ions that were significantly associated to patient survival in sarcoma patients: m/z (P =0.006) and m/z (P =0.022), both indicative of a poor overall survival in soft tissue sarcoma patients; and m/z (P = 0.017) indicative of a poor metastasis-free survival in myxofibrosarcoma patients. The Kaplan-Meier plots are shown in Fig.5.2. In order to increase the sample size for the subgroups, leiomyosarcoma, myxofibrosarcoma and undifferentiated pleomorphic sarcoma were grouped together as a single non-osteosarcoma (non-os) group, because these three subtypes are all soft tissue sarcomas, whereas osteosarcoma is primarily located in bone. Common 100

108 Chapter 5 Prognostic Metabolites biomarkers are beneficial from both a clinical and logistical viewpoint: for rare and sometimes diagnostically challenging tumors such as high grade sarcomas, which only present 1% of all malignancies and have over 50 histological subtypes, a prognostic biomarker common to many/all subtypes would be more broadly applicable, even for cases in which a definitive diagnosis is uncertain. Fig.5.2. Kaplan-Meier survival plots of prognostic biomarkers discovered from MALDI-TOF MSI datasets. Prognostic metabolite ions were found from a sub-dataset of myxofibrosarcomas and from a dataset consisting of all soft tissue sarcomas (non-os subset). (a) shows a metabolite ion (m/z ) that was found when the analysis was limited to undifferentiated areas, and was indicative for poor survival of soft tissue sarcoma patients. (b) and (c) show two metabolite ions (m/z and m/z ) indicating poor survival in soft tissue sarcoma patients and myxofibrosarcoma patients respectively. P values listed are calculated using a log-rank test in Kaplan-Meier analysis. The prognostic value of the metabolite ions were confirmed by visualizing the metabolite MSI data in patient samples with good and poor survival. Fig.5.3 shows the differential detection level of the metabolite ion at m/z in two non-os patients with differential survival. The metabolite MALDI MSI datasets used for the statistical analysis were obtained using a MALDI-TOF mass spectrometer. The mass accuracy and mass resolution of the system does not allow for the unambiguous assignment of the prognostic metabolite ions. Several poor survival and good survival patient tissues were reanalyzed using a high field FTICR mass spectrometer, in order to obtain high mass resolution high mass accuracy MALDI MSI data. Despite significant differences in mass analyzer technology (the time scale of detection of the MALDI-TOF instrument is approximately 10 5 times shorter than that of the FTICR) the mass spectral patterns were consistent, Supplementary Fig.5.4. As an example, the MS 101

109 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging images from MALDI-TOF and MALDI-FTICR for m/z are shown in Fig.5.4. The mass accuracy of the FTICR experiments was obtained using the metabolites ATP, ADP and AMP (images and experimental masses reported in Supplementary Fig.5.2), and was found to be less than 3 ppm (Note: the MSI dataset was not aligned nor recalibrated). Accordingly 3 ppm was the used for the maximum mass tolerance for querying metabolomics databases. It is now well established that MALDI MSI, using a given matrix preparation, detects a consistent subset of metabolites. Metabolite identities were assigned on the basis of accurate mass, and matched isotope distributions measured with a high field FTICR system to previously reported metabolite ions. MS/MS was attempted but the low signal intensities of the metabolite peaks and the presence of isobaric ions (see supplementary Fig.5.3) led to very noisy MS/MS spectra. Fig.5.3. MALDI-TOF MSI data of the prognostic metabolite ion at m/z (a) shows a comparison of the average spectra of two soft tissue sarcoma samples, one with good prognosis (blue line) and one with poor prognosis (red line). (b) shows the magnification of m/z Histological images and MSI visualizations for these two samples are shown in (c), confirming the higher detected intensity throughout the poor survival patient sample. 102

110 Chapter 5 Prognostic Metabolites Fig.5.4. Comparison of MALDI MSI images recorded using the MALDI-TOF AND MALDI- FTICR mass spectrometers. Histologic images of four samples are shown in (a-d). (e-h) are the MS images of inositol cyclic phosphate obtained with the MALDI-FTICR, and (i-l) the MS images obtained with the MALDI-TOF instrument. Within 2 ppm difference, the molecule at m/z was previously reported by Buck et al. at m/z in colon cancer and assigned as carnitine 173. The molecule observed at m/z was assigned as inositol cyclic phosphate using the Human Metabolome Database with an error of 0.32 ppm and matching the isotope pattern (Supplementary Fig.5.3). The ion detected at m/z could not be assigned using the available databases and MALDI-MSI literature. Here we detected the abundant metabolites that were statistically associated to patient survival. Metabolites signals that could differentiate between the different histological subtypes were also assessed, but none were statistically significant. Carnitine is an amino acid derivate involved in lipid metabolism and is known to contribute to multiple diseases including cancer 180. Carnitine plays an essential role in transporting fatty acids to mitochondria with the help of carnitine palmitoyltransferase I (CPTІ) for energy production by fatty acid oxidation (FAO) 181. FAO and CPTІ have been reported as emerging therapeutic targets in cancer 182. Inositol (1,2-) cyclic phosphate belongs to the class of organic compounds known as cyclitols. It can be transformed to inositol 1-phosphate (IP), which is involved in inositol 1,4,5-trisphosphate and calcium (IP3/Ca2+) signaling system 183. Altered expression of inositol IP3 receptor has been reported in different types of cancer and implicated in patient prognosis 184,

111 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging The prognostic metabolite biomarker inositol cyclic phosphate was found to be statistically significant only when the analysis was focused on the undifferentiated regions of soft tissue sarcoma samples. Undifferentiated tumor cells are abnormal-looking cells compared to moderately differentiated or well differentiated tumor cells, and are considered to be more malignant. The increased statistical significance observed when the analysis was focused on undifferentiated tumor areas is consistent with molecular and/or measurement variance being introduced by histological heterogeneity; the increased variance in the measured data undermines the ability to obtain statistically significant results. The work reported here could be expanded in several different directions: i) validate the results in a larger independent series; ii) assess whether inositol (1,2-) cyclic phosphate and carnitine are also prognostic for additional subtypes of softtissue sarcoma; iii) investigate the biological foundations of their prognostic value, which may ultimately lead to novel treatment strategies 186. Sample degradation can be a very important issue in metabolomics research, and several methods have been investigated for maintaining metabolic integrity of tissue samples for MALDI MSI 187,188. If the purpose of the study is to investigate metabolic processes then metabolic integrity is key. Maintaining metabolic integrity is a major challenge for clinical tissue samples, as those obtained via resection (the majority of those that are available) will have undergone degradation during surgery as blood vessels are sealed, especially considering it has previously been shown that post-mortem degradation occurs more rapidly at body temperatures than at room temperature 189. The more limited scope of a biomarker study, in which the goal is to find mass spectral features that differentiate between patient groups, in this instance between poor and good survival, does not require maintenance of the system s original metabolic state, only that the mass spectral features consistently differentiate between the groups. This work and previous work using MALDI of FFPE tissues 173, and DESI based analyses 174,175, have clearly demonstrated the utility of metabolic signatures for differentiating between patient groups. For clinical research in which metabolic integrity needs to be kept as close as possible to the original physiological state flash frozen needle biopsies are recommended. 5.4 Conclusions A prognostic metabolite biomarker discovery platform was setup for histologically 104

112 Chapter 5 Prognostic Metabolites heterogeneous high-grade sarcomas using fresh frozen tumor tissues. Inositol cyclic phosphate and carnitine were found to be potential generic prognostic biomarkers in soft tissue sarcoma patients. 5.5 Acknowledgments The authors thank Dr. C. Esteve for experiment assistance in MALDI-FTICR platform and thank B.E.W.M. van den Akker for her technical assistance. The authors would like to acknowledge financial support from COMMIT, Cyttron II and the ZonMW Zenith project Imaging Mass Spectrometry-Based Molecular Histology: Differentiation and Characterization of Clinically Challenging Soft Tissue Sarcomas (No ). BB is funded by the Marie Curie Action of the European Union (SITH FP7-PEOPLE-2012-IEF No ). 105

113 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Supplementary information Supplementary Fig.5.1. Data distribution of m/z Supplementary Fig.5.2. MALDI-FTICR-MSI experimental data of AMP, ADP and ATP. Note datasets not recalibrated (improved mass accuracy could be obtained by recalibrating on AMP, ADP and ATP). 106

114 Chapter 5 Prognostic Metabolites Supplementary information Supplementary Fig.5.3. Isotope pattern confirmation of peak assignments. 107

115 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Supplementary Fig.5.4. Comparison of MALDI-ToF and MALDI-FTICR spectra obtained from sequential tissue sections. a) and b) show the total mass spectra obtained from the 108

116 Chapter 5 Prognostic Metabolites Supplementary information MALDI-ToF and MALDI-FTICR respectively. c) h) show close ups of the regions of the prognostic metabolites reported here (indicated with colored bars). Supplementary Fig.5.5. Example MS images with and without TIC normalization. (a-c)/(g-i) are TIC normalized MS images (indicated with a T); (d-f)/(j-l) are the original images prior to TIC normalization, and which showed similar visualizations. The images are of the prognostic metabolite ions reported here, namely m/z , m/z and m/z

117 Biomarker Discovery in High Grade Sarcomas by Mass Spectrometry Imaging Supplementary Fig.5.6. Overview of the five metabolite ions detected by SAM analysis with FDR <5%. Upon Kaplan-Meier analysis two of the ions had P values greater than 0.05 (marked in red). Only those metabolite ions that were found to be associated with survival by SAM and Kaplan-Meier analysis were included in the main manuscript. 110