Diagnosis of prostate cancer by desorption electrospray ionization mass spectrometric imaging of small metabolites and lipids

Metabolite/Lipid Imaging of Prostate Tissue Specimens.

To investigate the lipidomic and metabolic profiles in normal and malignant prostate tissues, we selected 54 fresh frozen prostate tissue samples harvested at the time of radical prostatectomy and performed negative ion mode DESI-MSI. For each case, a 5-μm frozen section was taken for histopathological evaluation. After hematoxylin and eosin (H&E) staining, slides were annotated by a genitourinary pathologist to delineate areas of cancerous and normal tissue. A 15-μm section taken immediately adjacent to the H&E section was used for DESI-MSI. The details of DESI-MSI are given in Materials and Methods and depicted in SI Appendix, Fig. S1.

Metabolite and lipid ions were detected in the m/z range 50–1000. Fig. 1A shows a representative negative ion mode DESI mass spectrum of a typical prostate tissue specimen in the m/z range 50–200, where most of the small metabolites and glucose were detected. SI Appendix, Fig. S2 shows a representative image in the m/z range 200–1,000 where most lipids were detected. We identified and characterized several ion signals from the data (SI Appendix, Table S1) by using high mass accuracy, isotopic distribution, and tandem mass spectrometry. Typical molecular characterization data are presented in SI Appendix, Figs. S3 and S4 A–P. The detected species (SI Appendix, Table S1) were mostly deprotonated small metabolites related to energy production in the Krebs cycle, and deprotonated lipids including free fatty acids (FAs), FA dimers, phosphatidic acids, and glycerophospholipids. Using pixel-to-pixel mass spectral data, detailed 2D molecular images of the tissue were constructed to visualize the spatial distribution of selected individual metabolites and lipids. The spatial resolution of the image is ∼200 μm, which compares well with the thickness of a surgical knife. For example, Fig. 1B and SI Appendix, Fig. S2C show ion images of typical small metabolites and lipids, respectively, some of which appear to discriminate cancer from normal based on their relative abundance. The spatial distribution of many additional intense ion signals (m/z 89.0244, 303.2316, 423.2496, 585.4859, etc.) did not appear to be useful in differentiating normal from cancerous tissue (Fig. 1 and SI Appendix, Fig. S2).

Selection of Lipids for PCa Diagnosis.

Previous DESI-MSI studies have identified lipidomic profiles that distinguish normal from malignant tissues (16, 20, 21). We also attempted to distinguish cancerous from normal prostate tissues based on the abundance of specific lipid species. Given the large number of lipid ions (m/z 200–1,000) interrogated, we first tried an unbiased statistical approach, Lasso (26), to select a parsimonious set of lipids that could classify normal from cancerous tissues (SI Appendix, Fig. S5) (27). Using a training set of 45 specimens (17 normal and 28 cancerous), the Lasso considered m/z values that best characterized the two classes (cancer vs. normal) and yielded a cross-validation error of nearly 29% on pixel-based predictions (SI Appendix, Fig. S5C) and 27% on patient-based predictions. The lipid-based Lasso model did not work well when applied to the validation set, possibly because of the large heterogeneity in lipidomic profiles of PCa patients. Consistent with this finding, we were unable to validate previous work (28) identifying cholesterol sulfate as a candidate lipid biomarker for PCa (SI Appendix, Fig. S6).

Selection of Small Metabolites with Diagnostic Features.

We unambiguously identified and characterized a large number of small metabolites in the m/z window 50–200 that were differentially present in cancerous vs. normal tissue (Fig. 1A and SI Appendix, Table S1). For example, Fig. 1B shows higher abundance of erythrose (m/z 119.0348), glutamate (m/z 146.0464), glucose (m/z 179.0561), and sorbitol (m/z 181.0712), and lower abundance of succinate (m/z 117.0196), malate (m/z 133.0144), salicylate (m/z 137.0240), aconitate (m/z 173.0092), and citrate (m/z 191.0196) in cancerous tissues compared with normal. Most of these metabolites are important components of the Krebs cycle (Fig. 2). It should be noted that we could not construct DESI images of ionic species with very low abundances, even though they are important metabolites of the Krebs cycle and other metabolic pathways (see SI Appendix, Fig. S7 as typical example). Although imaging the spatial distribution and characterization of individual ions (metabolites) is not necessary for tissue diagnosis by DESI-MSI/Lasso, doing so can pinpoint important metabolites and their relation to PCa biochemistry (vide infra).

Lasso Analysis for the Training and Validation Sets.

We used Lasso to develop a classifier that estimates the probability that a pixel in each DESI-MS image is either cancer or normal. DESI-MS was performed on a training set of 36 tissue samples with histologically demarcated areas of normal and malignant tissue (18 pure normal and 18 pure cancer tissues from 36 patients). We selected the top 54 peaks of metabolites, which had been mostly characterized (SI Appendix, Tables S1 and S2), to perform the Lasso analysis. Two independent Lasso classifiers were built, one based on individual ion signals, and another based on ratios of different ion signals. The first classifier performed poorly in the validation set, perhaps caused by normalization issues. However, the second approach showed promising results in classification as detailed next. We divided the data in two m/z sets for Lasso: one with m/z range 50–1,000, which considers all detected metabolites and lipids, and the other with the m/z range 50–200, which considers only the small metabolites (SI Appendix, Table S3). Table 1 shows the patient-based prediction results from these calculations (see Statistical Analysis in Materials and Methods for details). The cross-validation analysis of all samples from the training set achieved nearly 92% overall agreement in the m/z range 50–1,000, and nearly 89% overall agreement in the m/z range 50–200, compared with standard histopathologic evaluation (H&E).

We also analyzed 18 independent validation specimens (10 normal and 8 cancer tissues) to test the performance of our Lasso classifier. Table 1 also shows the Lasso prediction obtained from the validation set in both m/z domains, per patients. Compared with histopathology (H&E), nearly 89% and 94% overall agreements were achieved in the m/z ranges 50–1,000 and 50–200, respectively. All these data collectively suggest that analysis of small metabolites can exhibit high accuracy in cancer diagnosis by DESI-MSI.

Glucose/Citrate Ratio as a Biomarker.

We also investigated whether we could preserve the good accuracy of the above analysis using fewer metabolites to make the testing more rapid. As discussed above, DESI-MSI reveals in situ metabolic flux by detecting the relative abundance of numerous metabolites. For example, Fig. 1B shows high levels of glucose (m/z 179.0561; see SI Appendix, Fig. S4I for identification) and low levels of citrate (m/z 191.0196; see SI Appendix, Fig. S4J for identification) in cancer (29). Decreased levels of citrate in PCa are well described and can also be detected using magnetic resonance (MR) spectroscopic imaging (30). Based on these findings, we attempted to use imaging of glucose and citrate (Fig. 1B) to characterize whether a prostate tissue specimen was cancerous or normal. Interestingly, when we calculated the ratio of these two ion signals (glucose/citrate) from the tissue specimen, and constructed the image of that ratio, we accurately differentiated PCa from benign prostate tissue as shown in Fig. 3. Furthermore, this specific ratio can also be chosen from the Lasso model (SI Appendix, Table S3), which is a linear function of log ratios and there are different but nearly equivalent ways of expressing the same fitted model.

In several malignancies including PCa, molecular changes have been reported in the normal tissues immediately adjacent to the cancer (field defects) and this could adversely impact detection of cancer by blurring the border between malignant and normal tissues (31). To determine whether there was a distinct border of the relative quantities of glucose and citrate between normal and malignant prostate tissues, we set our DESI probe to scan along a single line (32) over a prostate tissue specimen (Fig. 4). Among hundreds of ions recorded on the line scan, we extracted the signal intensities of glucose and citrate to plot them over the scan time (tissue length). Fig. 4 shows that there is a distinct cutoff at the border between normal and malignant prostate tissues as determined by the glucose/citrate ratio. This difference appears to be determined largely by citrate levels as its signal intensity drastically decreases in cancer, whereas the increase of glucose signal intensity was less marked. This difference could be caused by the fact that it is more difficult to ionize glucose than citrate in the gas phase by electrospray (SI Appendix, Fig. S8). Ion images of glucose or citrate alone did not work so well as the ratio between the two metabolites (Fig. 3) in identifying PCa. It should be noted that inflammation is often observed in a PCa specimen (Fig. 3B) and this glucose/citrate ratio fails to distinguish it. A follow-up study is required in the future for finding biomarkers that can distinguish inflammation from malignant or normal tissue.

To test the ability of the glucose/citrate ratio to diagnose PCa, we used the same training set that was used above for the Lasso, comprising 36 specimens. We evaluated the average glucose/citrate signal ratio from each individual specimen for the cancerous and normal areas based on histology on the adjacent H&E slide. There was a significant difference in the distribution of the glucose/citrate ratios between PCa and normal tissues, and specimens could be accurately classified as cancer if the glucose/citrate signal ratio was >1 and normal if the ratio was <0.5 (Fig. 5A). In an independent set of 18 specimens (8 cancer and 10 normal tissues) we were able to validate that all specimens with glucose/citrate signal ratio >1 were malignant and all with ratios <0.5 were normal (Fig. 5B). There were nine specimens in the training set (Fig. 5A), and two specimens in the validation set (Fig. 5B) with a ratio between 0.5 and 1; the significance of this ratio is unknown, although some of those were classified as normal and some as cancer by H&E.

Preparation of Prostate Tissue Specimens.

Following Institutional Review Board approval, we accessed fresh-frozen prostate tissue obtained from prostatectomy specimens and stored in the Stanford University Department of Urology tissue bank. The samples included 64 prostate tissue specimens (28 benign only, 36 cancer only, and 10 cancer mixed with benign tissue) that had been obtained from different patients and stored at −80 °C until sectioning. A 5-μm section of each tissue sample was mounted on a glass slide and stained with H&E. A genitourinary pathologist (C.A.K.) reviewed and annotated normal and cancerous areas on the H&E-stained 5-µm adjacent tissue sections. Benign tissue included areas of normal prostate and areas of inflammation. 15-µm-thick adjacent sections were obtained using a Leica CM1950 cryostat (Leica Biosystems). These tissue sections were mounted on glass microscope slides and stored at −80 °C before DESI-MS analysis. The 54 specimens were used to construct a training set (n = 36) composed of 18 normal and 18 cancer (Gleason pattern 3 and/or Gleason pattern 4) specimens, and a validation set (n = 18) composed of 10 normal and 8 cancerous specimens for DESI-MS evaluation. The remaining 10 specimens of mixed grade were evaluated by the distribution of the glucose/citrate ion signal ratio to determine cancer margins.

DESI-MSI Study.

The detailed method of tissue imaging by DESI has been described elsewhere (21, 42). Briefly, a laboratory-built DESI source coupled to an LTQ-Orbitrap XL mass spectrometer (Thermo Scientific) was used for tissue imaging. DESI-MSI was performed in the negative ion mode (−5 kV) from m/z 50–1,000 with a spatial resolution of 200 µm (spray spot diameter) using a histologically compatible solvent system 1:1 (vol/vol) dimethylformamide/acetonitrile (DMF/ACN) (43) at a flow rate of 0.7 µL/min. The HPLC-grade solvents (DMF, ACN, etc.) were purchased from Sigma-Aldrich. Nitrogen, at a pressure of 170 psi, was used as the sheath gas for electrospray nebulization. The prostatic tissues were scanned under impinging charged droplets using a 2D moving stage in horizontal rows separated by a 200-µm (spatial resolution) vertical step. All imaging experiments were carried out under identical experimental conditions including geometrical parameters, e.g., spray tip-to-surface distance ∼2 mm, spray incident angle of 55°, and spray-to-inlet distance ∼5 mm. Data acquisition was performed using XCalibur 2.2 software (Thermo Fisher Scientific Inc.). An in-house program allowed the conversion of the XCalibur 2.2 mass spectral files (.raw) into the image file, which could be read by a biomedical image analysis software called Biomap (freeware, https://ms-imaging.org/wp/. The distributions of different metabolites, lipids, and the glucose/citrate ratio were plotted (Figs. 1B and 3, and SI Appendix, Fig. S2C) using the Biomap software. In ion images, we have used rainbow color order to represent the highest concentration by red and the lowest concentration by violet.

Metabolite/Lipid Identification.

The ions observed in the DESI-MS study were identified by searching the MassBank (www.massbank.jp) and the LIPID Metabolites and Pathways Strategy (www.lipidmaps.org/) databases based on high mass accuracy and isotopic distribution. When the database listed multiple isobaric/isomeric metabolites or lipids, we performed collision-induced dissociation (CID) and compared the corresponding fragmentation profile with that of the standard from the above database to characterize the species (44) wherever applicable (see SI Appendix, Fig. S6 A–P as typical examples). The CID spectra of the mass selected ions from the PCa tissue specimen are sometimes complex, although the majority of fragment ions matches with that of standards. This complexity can be interpreted by the interference of isomeric/isobaric ions derived from the biological matrix (tissue). As the position and stereochemistry of the double bond in an FA complicates the structural elucidation, they are often tentatively assigned in FAs and glycerophospholipids. However, some common and well-known metabolites and lipids that are commercially available were identified from the tissue by comparing their MS/MS data with standards.

Post-DESI-MSI Histopathological Reevaluation.

As the DESI solvent DMF/ACN (1:1; vol/vol) is histologically compatible (43), the same tissue samples, after analyzing in DESI-MS, were subjected to H&E staining followed by histopathological reevaluation. The dyes and solvents used in H&E staining were purchased from Fisher Scientific. High-resolution (0.43-µm) optical images of H&E-stained tissues were recorded using a Hamamatsu NanoZoomer 2.0-RS slide scanner. The pathology examination was made without the knowledge of DESI-MS evaluation.

Statistical Analysis.

XCalibur raw data files were converted to .txt files for statistical analysis. The raw data in .txt file format were imported to the R language for statistical analysis. Although hundreds of metabolites and lipids were detected by DESI-MS, we selected the top 54 peaks (SI Appendix, Tables S1 and S2), whose abundances were significant, and most of them (SI Appendix, Table S1) were characterized by tandem mass spectrometry. We performed statistical analysis both by using the individual peaks and by using all possible ratios of two peaks in this list (SI Appendix, Tables S1 and S2). Within the training set (18 benign, and 18 cancer), we applied the Lasso method (multiclass-logistic regression with L1 penalty) using the glmnet package in the CRAN R language library (45).

The Lasso yields sparse models, that is, models that involve only a subset of the variables/predictors (27). Therefore, models generated using the Lasso are simpler and easier to interpret than those from other linear regression methods. In our application, the Lasso method yields a model with parsimonious sets of features for discriminating cancer and normal prostate tissue. A mathematical weight for each statistically informative feature is calculated by the Lasso depending on the importance that the mass-spectral feature has in characterizing a certain class. Because the features selected by the Lasso can occur at a valley or a shoulder of an actual mass-spectra peak, identification of the selected features was performed by characterizing the nearest mass-spectra peak to the statistically selected feature. Classification was done on a pixel-by-pixel basis into one of two classes: cancer or normal, and then results were converted to patientwise predictions using a majority rule: if the majority of a tissue (patient) pixels are predicted to be cancer, the tissue is predicted as cancerous. We used cross-validation, leaving out one patient at a time, to select the Lasso tuning parameter and to assess the predictive accuracy within the training set. Then, the chosen model was applied to the test set of 18 patients.

Additional Information.

Supporting information accompanies this paper in the SI Appendix.