Spatially Resolved Metabolomic Characterization of Muscle Invasive Bladder Cancer by Mass Spectrometry Imaging

Abstract

Muscle invasive bladder cancer (MIBC) is an advanced stage of bladder cancer which poses a severe threat to life. Cancer development is usually accompanied by remarkable alterations in cell metabolism, and hence deep insights into MIBC at the metabolomic level can facilitate the understanding of the biochemical mechanisms involved in the cancer development and progression. In this proof-of-concept study, the optimal cutting temperature (OCT)-embedded MIBC samples were first washed with pure water to remove the polymer compounds which could cause severe signal suppression during mass spectrometry. Further, the tissue sections were analyzed by infrared matrix-assisted laser desorption electrospray ionization mass spectrometry imaging (IR-MALDESI MSI), providing an overview on the spatially resolved metabolomic profiles, which enabled the discrimination between not only the cancerous and normal tissues, but also the subregions within a tissue section associated with different disease states. Using t-Distributed Stochastic Neighbor Embedding (t-SNE), the hyperdimensional MSI data was mapped into a two-dimensional space to visualize the spectral similarity, providing evidence that metabolomic alterations might have occurred outside the histopathological tumor border. Least absolute shrinkage and selection operator (LASSO) was further employed to classify sample pathology in a pixel-wise manner, yielding excellent prediction sensitivity and specificity up to 96% based on the statistically characteristic spectral features. The results demonstrate great promise of IR-MALDESI MSI to identify molecular changes derived from cancer and unveil tumor heterogeneity, which can potentially promote the discovery of clinically relevant biomarkers and allow for applications in precision medicine.

Introduction

Bladder cancer ranks the 4th most common cancer among men in the United States (www.urologyhealth.org). Approximately 25% of bladder cancer cases are muscle invasive bladder cancer (MIBC) with an overall 5-year survival rate of 40–60% after being treated by radical cystectomy (Ogawa et al. 2020) and/or chemo-radiation in the adjuvant or neoadjuvant setting (Ma and Black 2021). Many studies have shown that metabolomes and lipidomes undergo significant changes during cancer development to allow rapidly growing cancer cells to maintain viability (Lin et al. 2017). A thorough understanding of cancer metabolism can offer invaluable insights into key biological events involved in cancer initiation and progression, and aid in identifying diagnostic biomarkers (Morse et al. 2019). Due to the enormous diversity in metabolite structures and characteristics, the metabolomic and lipidomic analysis of cancer primarily relies on chromatography-mass spectrometry (Cui et al. 2018; Struck-Lewicka et al. 2015; Urayama et al. 2010) due to its high molecular specificity and coverage, in which case generally a big piece of specimen is homogenized and undergoes extensive sample preparation where tumor heterogeneity is neglected. Herein, an imaging approach retaining spatial information so allowing assessments of microenvironments is desirable. Imaging of proteins in biological tissues is well established using labeling techniques, e.g, immunohistochemistry to map target antigens, while spatial detection of small metabolites is still under development. Over the last decade, innovative imaging modalities such as magnetic resonance imaging (MRI) have been exploited in clinical laboratories as the non-invasive nature enables their use in monitoring cancer metabolism in vivo (Lin et al. 2017). However, these modalities are subject to several inherent limitations such as narrow chemical information, poor sensitivity and specificity, and requirement of labeling.

Mass spectrometry imaging (MSI) is an objective imaging method of which feasibility of assessing microenvironments within tissue sections has been proven on a broad variety of diseases lately (Banerjee et al. 2017; Calligaris et al. 2014; Dória et al. 2016a; Jarmusch et al. 2016). The label-free nature and qualitative/ quantitative capability of MSI allow for the investigation of spatial distributions of numerous biomolecules simultaneously, which is well suited for untargeted discovery studies. Ion images are constructed by incorporating the spatial locations and ion abundances, which can be associated with the corresponding histological images to study the biomolecular signatures from specific subregions at distinct disease states. A substantial amount of recent cancer studies with desorption electrospray ionization (DESI) have demonstrated the value of ambient MSI in depth molecular profiling, and accurate cancer diagnosis with minimal sample preparation and analysis in near-physiological conditions (Cordeiro et al. 2020; Eberlin et al. 2014; Margulis et al. 2018; Pirro et al. 2017). Another emerging ambient MSI technique is infrared matrix-assisted laser desorption electrospray ionization (IR-MALDESI) which uses a mid IR-laser to resonantly excite water for direct tissue characterization. The complete ablation of a tissue slice over the irradiated area owing to the micrometer-scale penetration depth of an IR laser enables reproducible volumes to be sampled. The ionization mechanism of IR-MALDESI is demonstrated to be the post-electrospray ionization of ablated neutral materials (Dixon and Muddiman 2010; Rosen et al. 2015; Sampson et al. 2008; Tu and Muddiman 2019a). Over the past decade, IR-MALDESI MSI has established its utility for studying the lipid and metabolite distributions in various subjects including but not limited to cancerous ovaries (Nazari and Muddiman 2016), single mammalian cells (Xi et al. 2020) and fresh bones (Khodjaniyazova et al. 2019), with high repeatability (Tu and Muddiman 2019b) and good lipidome and metabolome coverage (Pace et al. 2020), indicating its high potential in cancer research.

One challenge in analyzing clinically relevant tissue is that the frozen tissue sample is oftentimes preserved in an embedding medium to assist cryosectioning and maintain its morphological characteristics. The most common medium used nowadays is optimal cutting temperature (OCT) which consists of polyethylene glycol, polyvinyl alcohol and nonreactive ingredients (Snijders et al. 2019; Vrana et al. 2016). Nonetheless, the polymers from OCT raise difficulties in MS analysis as they severely suppress informative biological signals by competing for the available ions. In this proof-of-concept study, we demonstrated a simple and effective washing step to remove OCT from tissue samples without impacts on the tissue morphology. The washed samples were then subject to positive and negative IR-MALDESI MSI for characterization of the metabolomic and lipidomic profiles of MIBC. Multivariate statistical analyses were further conducted in an attempt to unravel biomolecules associated with disease states and perform cancer diagnosis relying on the m/z patterns.

Experimental

Human Bladder Tissue Collection

Deidentified human bladder tissues including 6 tumor samples with their 6 adjacent matching normal tissues were collected at Tumor Tissue and Pathology Shared Resources (TTPSR) at Wake Forest Baptist Medical Center (WFBMC) and Comprehensive Cancer Center under approved IRB (IRB00036014). The tissue samples were snap-frozen, embedded in OCT blocks and then stored at −80 °C until analysis. The demographic and clinical information of all tissue samples is detailed in supplementary Table S1.

Histology

All tumor specimens were high grade MIBC that showed significant loss of urothelial morphology and deep infiltration and invasion of detrusor muscle. Tumors were of stage pT2b, pT3 and pT4 according to The 2016 WHO Classification of Tumors of the Urinary System (Humphrey et al. 2016) (Table S1). Adjacent histologically non-malignant “normal” tissues mainly comprised stroma and smooth muscles.

Sample Preparation

Human bladder tissues embedded in OCT media were cryosectioned into 20-μm sections using a cryomicrotome (CM 1950, Leica Biosystems, Buffalo Grove, IL, USA) at −20 °C, and thaw-mounted onto microscope slides. To remove OCT, each sample slide was submerged into pure water for 60 s, followed by into fresh water for another 30 s with gentle and careful stir during the washing steps. Sample slides were allowed to air dry for approximately 10 min prior to imaging.

IR-MALDESI MSI

IR-MALDESI employs a “burst-mode” 2970-nm laser with a pulse rate up to 10 kHz (JGM Associates, Inc., Burlington, MA) to induce sample ablation. 6 pulses/burst with a total energy of approximately 1.1 mJ was applied in each laser shot. An ice layer was uniformly deposited on the sample surface as a mid-IR absorbing matrix by cooling the sample stage down to −8 °C and keeping the relative humidity in the sample enclosure at approximately 12% (Robichaud et al. 2014). The electrospray solvent of 50:50:0.2 v/v/v acetonitrile/water/formic acid (Bagley et al. 2020) was infused at a flow rate of 1.5 μLmin⁻¹ for post-electrospray ionization of ablated neutral species. All solvents were in Optima grade and purchased from Fisher Scientific (Fair Lawn, NJ, USA).

The IR-MALDESI source is coupled to an Orbitrap Exploris^™ 240 mass spectrometer (Thermo Fisher Scientific, Bremen, Germany). The mass spectrometer was operated in “Small Molecules” mode with a RF lens of 70%. The spray voltage was set to be 3.5 kV and 3.2 kV for positive and negative mode, respectively. The capillary temperature was 350 °C. Mass spectra of all samples were acquired over m/z range 160–1000, with a resolving power of 240,000_fwhm at m/z 200. Mass calibration was performed daily prior to analysis using the Pierce^™ FlexMix^™ Calibration Solution (Thermo Fisher Scientific), and the EASY-IC^™ ion source was on for real-time mass calibration. The automatic gain control function (AGC) was disabled, and the injection time was held constantly as 15 ms to coordinate the laser desorption and ion acquisition events. For each tissue sample, half was subjected to positive IR-MALDESI MSI and the other half was to negative IR-MALDESI MSI. Tissue sections were imaged separately, generating a total of 12 positive and 12 negative MSI data sets. Pixel size was 100 μm-by-100 μm.

Data Processing and Visualization

The initial data files were stored in Xcalibur raw file format and converted into mzML files using MSConvert from the ProteoWizard toolkit, and then converted into imzML image files by imzMLConverter. imzML files were loaded into MSiReader v1.02 (https://msireader.wordpress.ncsu.edu/) for visualization, preprocessing and analysis (Bokhart et al. 2018; Robichaud et al. 2013). Ion images were presented with ± 2.5 ppm mass tolerance. The MSiPeakfinder was used to select species that displayed more than 2-fold abundances in either tumors or normal tissues in comparison with the background area and were above an abundance threshold of 100. Consecutively, 1338 and 685 ions in positive and negative mode were selected as tissue-specific, respectively. The MSiReader built-in MSiExport tool was used to export ion abundance values from each pixel for further statistical analyses.

Metabolites and lipids were putatively annotated via METASPACE (Palmer et al. 2017) using the HMDB-v4 and LipidMaps-2017–12-12 with a false discovery rate of 10%. Only ions marked as “on sample” were reported. The imaging files are public at https://metaspace2020.eu/project/tu-2021, and the annotation results are available in Supplemental File 1. METLIN (Guijas et al. 2018) was alternatively used to annotate monoisotopic masses (< ±2.5 ppm mass error) that were not annotated by METASPACE, such as ions detected as [M+H⁺-H₂O]⁺.

Multivariate Statistical Analyses

t-Distributed Stochastic Neighbor Embedding (t-SNE) is an unsupervised and non-linear dimensionality reduction algorithm that has been introduced to the MSI field recently (Abdelmoula et al. 2016; Fonville et al. 2013). In this study it was utilized to visualize the high-dimensional MSI data in a two-dimensional space to expose the similarity relationships of data points from different disease states. Tissue-specific ion abundances were exported from subregions corresponding to varying pathologies, i.e., 100 pixels on each sample (either tumor or normal), and 100 pixels on the adjacent smooth muscle regions in tumor #1999–129 and #2002–322. The subregions were reviewed and annotated through histopathological evaluation of the H&E stained slides, which are shown in Fig. S1 and Fig. S2 together with the pixels of interest used for statistical analysis. Data was first normalized to the total ion current (TIC) and log-transformed, and was processed by t-SNE with the MatLab function tsne using the default settings.

To demonstrate the feasibility of IR-MALDESI MSI to identify characteristic biomarkers and differentiate sample types, least absolute shrinkage and selection operator (LASSO) as an embedded method which simultaneously processes feature selection with a machine learning classifier was used in this study. LASSO is a linear regression model with the goal of minimizing the residual sum of squares of the model subject to the sum of the absolute values of the coefficients which has to be less than a fixed value (Muthukrishnan and Rohini 2016). To do so, it shrinks the coefficients of the less associated variables to 0 to remove the redundant variables and thus select the most discriminative ones for the subsequent prediction purpose. The LASSO model was trained on the mass spectral data from all samples except #2002–342 using the “glmnet” package v4.1 in the R language. A total of 1000 pixels from the five patients were merged in the training set, where 100 pixels were from each sample (either tumor or normal tissue). The regularization parameter λ was determined by 10-fold cross-validation on the training set, and the one at the minimum misclassification error was chosen for the LASSO model. The model was subsequently tested on an independent sample pair from Patient #2002–342 to predict the sample type (i.e., tumor or normal) in a pixel-wise manner, and the classification sensitivity (true positive rate) and specificity (true negative rate) were calculated to evaluate the predictive performance.

Results and Discussion

Molecular Profiling of Human Bladder Tissues

Removal of the OCT embedding media was necessary prior to the MS analysis to avoid signal suppression. Fig. S3 shows that after approximately 90 s of dipping the sample slide into pure water with gentle agitation, the OCT layer was removed and the polymer peaks were vastly reduced (e.g., average ion flux at m/z 476.6260 was 2.4×10⁵ versus 800 before/ after washing), while holding the slide stationary resulted in predominant OCT background which was indicative of incomplete removal of the OCT compounds.

After tissue washing, 6 matching pairs of bladder tissues were investigated by positive and negative IR-MALDESI MSI. Overall, dramatic spectral differences can be observed between tumor and normal tissues, while the adjacent smooth muscle regions in tumor show a certain extent of similarity to both tumor and normal (Fig. 1). The tumor samples appear to exhibit higher ion abundances in the mass range of m/z 650–850 in positive mode and m/z 250–350 in negative mode, which are typically associated with phospholipids (PLs) and long-chain fatty acids (FAs), respectively. The normal samples were found to contain more larger lipids in the mass range of m/z 850–1000 in positive mode, e.g., triglyceride TG (54:4) with NH₄⁺ adduct at m/z 900.8030.

Fig. 1.

Representative mass spectra recorded from bladder tumor (top), adjacent smooth muscle regions within tumor (middle) and normal tissue (bottom). Inserts: zoom-in spectra at m/z 550–1000.

In regard to the coverage of biomolecules, more species were putatively identified in the tumor samples compared to their normal counterparts. For instance, in positive mode 622 and 388 unique annotations were assigned to the tumor and to the control via METASPACE LipidMaps database, respectively, and in negative mode it was 435 versus 162 unique annotations (Supplementary Material 1). The vast majority of detected species in positive mode were classified as phosphatidylcholine (PC), phosphatidylethanolamine (PE), ceramide (Cer), sphingomyelin (SM), glycerides, steroid and cholesteryl ester (CE). In negative mode the detected classes are mainly FA, hydroxy FA, FA ester, phosphatidylglycerol (PG), phosphatidylinositol (PI), phosphatidylserine (PS) and ceramide phosphate. Certain molecules detected in positive mode were also present in negative mass spectral, such as monoglyceride (MG), diglyceride (DG) and PE as deprotonated species.

Visualization of Molecular Alterations between Tumor and Normal Bladder

Two-sample t-test was carried out on all tissue-specific ions (1338 in positive mode and 685 in negative mode; please find the peak selection criteria in Data Processing and Visualization) to determine the statistical significances (i.e., p-values) of their abundances in tumors relative to normal tissues. Then the p-values were presented against the log₂ fold changes in ion abundances via the volcano plot (Fig. 2a and 2b) to identify significant changes, where the upregulated species in tumor versus normal bladder tissues were located on the right and the downregulated species were on the left. The ions with p-values ≤0.01 and fold changes ≥2 were considered significantly altered in bladder cancer and putatively annotated through database searching, which are listed in Supplementary Material 2. The results revealed distinct lipid and metabolite constituents between tumor and normal tissues. In positive mode, PLs including lysoPC, PC, lysoPE and PE species were amongst the most abundant in tumors (Fig. 3 and Fig. S4), which corroborates the observations with the mass spectra. The overexpression of PC and PE composition has been frequently reported in diverse cancer types (Mirnezami et al. 2014; Ouyang et al. 2017; Porcari et al. 2018). PC and PE are the first two most common constituents of the lipid bilayer (Szlasa et al. 2020), participating in various cellular processes such as cell division and proliferation (Mao et al. 2016), bioactive molecules production (e.g., DG (Momchilova and Markovska 1999)), and membrane protein folding (Bogdanov and Dowhan 1998). Glycerides including MGs and DGs are also present at higher abundances in tumor.

Fig. 2.

Volcano plots of tissue-specific ions (−log₁₀ of the t-test p-value versus log₂ of the fold change tumor/normal) in (a) positive and (b) negative mode. Each data point in the volcano plots represents a tissue-specific ion of which selection criteria are detailed in Data Processing and Visualization. Representative statistically changed metabolites are labeled in the volcano plots. t-SNE scatter plots visualize the MSI data points from tumor (n=600, red filled circles), adjacent muscle in tumor (n=100, green filled circles) and normal tissue (n=600, blue filled circles) in (c) positive and (d) negative mode in two dimensions.

Fig. 3.

Selected ion images from the tumor (T) and normal (N) samples acquired by positive mode IR-MALDESI MSI: left, upregulated species; right, downregulated species. Putative annotations were obtained via METASPACE database search. The corresponding H&E-stained images of adjacent tissue sections are shown on top, where the smooth muscle regions were outlined with red lines, except for tumor #1999–129 where the tumor regions were outlined with blue lines and the rest area was smooth muscle. The scale bar represents 1 mm.

Nonetheless, not all PLs are elevated in tumor. For instance, the normal tissues are rich in very-long-chain PEs such as PE (52:4) (Fig. 3). In addition, some wax ester (WE), sphingomyelin (SM) and cholesteryl ester (CE) species display higher ion abundances in the normal tissues, such as SM (d36:1), SM (d42:3) and CE (15:0) (Fig. 3 and Fig. S5). The ion images also clearly reflected the intratumor heterogeneity. For instance, MG (18:1) showed a specific localization in the tumor regions but was weak or absent in the neighboring benign smooth muscle regions in the tumor samples, while an increased abundance of CE (15:0) behaves in the opposite manner.

In negative mode, one class of molecules observed with prominent and consistent increase in tumor was free fatty acids (FAs) (Fig. 4 and Fig. S6). This result is in accordance with the published studies on other cancer types (Mao et al. 2016; Margulis et al. 2018). FAs are the essential building blocks for a large variety of lipid species (Koundouros and Poulogiannis 2020). Elevated biosynthesis of de novo FAs in cancer cells not only maintains the rapid cellular proliferation (Baenke et al. 2013), but also serves as an essential energy source (Röhrig and Schulze 2016). Some complex lipid species such as PI and PA, were also found concentrated in the bladder tumors (Fig. S7). The level of PI expression has been reported to markedly increase in prostate cancer (Goto et al. 2014). This could be related to the elevated cellular signaling activity (Krauß and Haucke 2007). PA class was also found richer in ovarian cancer versus control (Dória et al. 2016b). The normal bladder tissues show an enhancement in some glutamine derivatives, such as γ-glutamyl-β-aminopropiononitrile at m/z 198.0884 (Fig. 4). This could potentially be linked to an elevated γ-glutamyl transferase activity (a clinical indicator of tissue damage (Heisterkamp et al. 2008)) in tumor, and γ-glutamyl-β-aminopropiononitrile is one of its substrates. The tumor heterogeneity was also exhibited in negative mode, allowing clear visualization of the tumor regions and the neighboring benign smooth muscle regions within the tumor samples.

Fig. 4.

Selected ion images from the tumor (T) and normal (N) samples acquired by negative mode IR-MALDESI MSI: left, upregulated species; right, downregulated species. Putative annotations were obtained via METASPACE database search. The corresponding H&E-stained images of adjacent tissue sections are shown on top, where the smooth muscle regions were outlined with red lines, except for tumor #1999–129 where the tumor regions were outlined with blue lines and the rest area was smooth muscle. The scale bar represents 1 mm.

Data analysis and interpretation in MSI studies can be difficult due to the high dimensionality of the data. Manual inspection of individual ion images to distinguish tumor can be time- and labor-consuming and tend to neglect the correlations between ions. t-SNE as a nonlinear dimensionality reduction method was performed on the spectral data from pathology-defined subregions of the bladder tissues in order to visualize the results of the high-dimensional MSI datasets in a low-dimensional representation while retaining a lot of the original information. Compared with linear methods which mainly focus on the global data structure, nonlinear dimensionality reduction approaches can preserve both local similarities and global characteristics (Abdelmoula et al. 2016). The two-dimensional scatter plots are displayed in Fig. 2c and 2d, yielding apparent separation of tumor and normal data points, indicating that their molecular signatures are characteristically different. The moderate dispersed grouping could mainly be attributed to the patient-to-patient variability. It is interesting to note that the data points from the muscle cells in the vicinity of tumor fell between the tumor and the normal clusters, likely indicating the neighboring smooth muscle regions have mixed molecular profiles of both tumor and normal tissues. For instance, a tumor-upregulated ion at m/z 263.2369, putatively identified as octadecatrienal and isomers, in adjacent smooth muscles showed the same level as in normal (p=0.26), whereas PE (40:2) at m/z 800.6164 in adjacent smooth muscles expressed similar level as in tumor (p=0.22), which was 15× less in normal (p=5.81×10⁻²³) (Fig. S8). Indeed, some other studies have also shown that the molecular tumor margin could exceed the histologically defined tumor margin (Banerjee et al. 2017; Oppenheimer et al. 2010), implying possible aberrant cell events occurring in the immediately adjacent normal cells. The hypotheses proposed by Oppenheimer et al (Oppenheimer et al. 2010) for this phenomenon were that the seemingly normal cells might be infiltrative tumor cells or precancerous cells; or it could be a result of the interactions between the tumor and adjacent cells. Our study provides additional evidence that the underlying molecular alterations might have already occurred in the normal-appearing regions beyond the histological tumor border, underlining the cellular complexity of tumor.

LASSO Prediction of Bladder Cancer and Corresponding Normal Tissue

A binary LASSO model was constructed to fulfill feature selection and disease classification using the training set including data points from five out of six patients. 10-fold cross-validation was employed for parameter tuning and the regularization parameter λ at the minimum misclassification error was chosen for the LASSO model. A total of 40 and 61 m/z features in positive and negative ionization mode, respectively, showing the best discriminative power were selected by LASSO and are displayed in Fig. 5a and Supplementary Material 3 with their putative identifications, where the features with positive weights represent that they are characteristic of tumor, while negative weights represent the opposite.

Fig. 5.

(a) Discriminating mass spectral features with their weights obtained by the LASSO model in positive and negative mode. Putative annotations were made via METASPACE, which were detailed in Supplementary Material 3. (b) LASSO per-pixel prediction results on Patient #2002–342.

To evaluate its performance for MIBC diagnosis based on the IR-MALDESI MSI data, the LASSO model was further applied to an independent dataset (i.e., patient #2002–342) to predict the sample class per pixel (Fig. 5b), yielding a sensitivity of 96.1% and a specificity of 91.1% in positive mode, and a sensitivity of 90.4% and a specificity of 96.3% in negative mode, indicating that the spectral features detected by IR-MALDESI are highly predictable.

Conclusions

This work demonstrates the feasibility of IR-MALDESI MSI in capturing informative molecular fingerprints that are indictive of disease states. In combination with a dimensionality reduction approach, cancerous and normal tissues were differentiated and potential molecular alterations beyond the histological tumor margin were revealed. LASSO as a feature selection technique and machine learning classifier allowed for identifying discriminative biomolecules and accurate prediction of disease states of new data. The results offer a proof of concept that IR-MALDESI MSI can be a complementary imaging approach to the existing imaging modalities for cancer research to provide in-depth molecular information and a comprehensive picture of cancer metabolism, which may provoke vital clinical implications. Note that the current work on the limited number of samples was exploratory and preliminary. In the future study the findings in this work will be validated in a larger patient cohort, and the significantly changed species will be unambiguously identified via tandem MS.

Abbreviations

PC

phosphatidylcholine

PE

phosphatidylethanolamine

PI

phosphatidylinositol

PS

phosphatidylserine

PG

phosphatidylglycerol

FA

fatty acid

MG

monoglyceride

DG

diglyceride

TG

triglyceride

CE

cholesteryl ester

Data Availability

The data sets are available at https://metaspace2020.eu/project/tu-2021.