The incidence of esophageal adenocarcinoma is rising, survival remains poor, and new tools to improve early diagnosis and precise treatment are needed. Cancer phospholipidomes quantified with mass spectrometry imaging (MSI) can support objective diagnosis in minutes using a routine frozen tissue section. However, whether MSI can objectively identify primary esophageal adenocarcinoma is currently unknown and represents a significant challenge, as this microenvironment is complex with phenotypically similar tissue-types. Here, we used desorption electrospray ionization-MSI (DESI-MSI) and bespoke chemometrics to assess the phospholipidomes of esophageal adenocarcinoma and relevant control tissues. Multivariate models derived from phospholipid profiles of 117 patients were highly discriminant for esophageal adenocarcinoma both in discovery (AUC = 0.97) and validation cohorts (AUC = 1). Among many other changes, esophageal adenocarcinoma samples were markedly enriched for polyunsaturated phosphatidylglycerols with longer acyl chains, with stepwise enrichment in premalignant tissues. Expression of fatty acid and glycerophospholipid synthesis genes was significantly upregulated, and characteristics of fatty acid acyls matched glycerophospholipid acyls. Mechanistically, silencing the carbon switch ACLY in esophageal adenocarcinoma cells shortened glycerophospholipid chains, linking de novo lipogenesis to the phospholipidome. Thus, DESI-MSI can objectively identify invasive esophageal adenocarcinoma from a number of premalignant tissues and unveils mechanisms of phospholipidomic reprogramming.

Significance:

These results call for accelerated diagnosis studies using DESI-MSI in the upper gastrointestinal endoscopy suite, as well as functional studies to determine how polyunsaturated phosphatidylglycerols contribute to esophageal carcinogenesis.

Esophageal cancer is the eighth most common malignancy worldwide causing approximately 500,000 deaths per annum (1). The predominant subtype in Western countries is esophageal adenocarcinoma (EA), which is linked to esophageal acid reflux. The normal lower esophageal squamous epithelium is replaced with an intestinal-type columnar mucosa (“Barrett metaplasia”), which can acquire dysplastic changes leading to EA (2). It is often detected at an advanced stage, with approximately 30% of patients amenable to curative treatment (3). Despite improvements in multi-modality treatment, 5-year survival after treatment with curative intent remains 30%–35%, and all-stage 5-year survival is 14% (3). Therefore, new tools to facilitate early diagnosis and effective treatment are needed.

Despite extensive investigation of how metabolism contributes to cancer (4–6), lipids, which constitute 70% of the human metabolome, are relatively understudied (7). Most lipids are sequestered in bilayer membranes as glycerophospholipids. There are six glycerophospholipid classes based on their head moiety, with further variation in acyl lengths and desaturations. These features convey diversity to the glycerophospholipid profile (8), which is often altered in cancer states (9–12). Their chemical and physical stability makes them attractive for biomarker studies. Functionally, glycerophospholipid diversity impacts membrane characteristics and cellular signaling. For example, the glycerophospholipids phosphatidylinositols (PI) mediate PI3K signaling, which is one of the most commonly deregulated pathways in EA (13), whereas phosphatidylglycerols have potent and contrasting effects on squamous cells' proliferation (14). Thus, measuring glycerophospholipids may provide diagnostic biomarkers with actionable therapeutic potential; however, species composition must be determined precisely.

Tissue phospholipidomics using traditional LC/MS was inherently limited, as even small homogenates cannot eliminate nontarget cell contamination in complex microenvironments like cancer (15, 16). Mass spectrometry imaging (MSI; ref. 17) addresses this, by providing spatially resolved lipid analysis using tissue sections. This allows accurate phospholipid profiles to be derived from histologically pure areas. Desorption electrospray ionization (DESI)-MSI offers advantages over MSI approaches, due to range of lipids detected, minimal sample preparation, performance in ambient conditions, and nondestruction of analyzed tissue sections allowing for comparative histologic analysis (18). Typically hundreds of pixels from several sampling zones per specimen are collected, such that each phospholipid profile comprises tens of thousands of mass spectra across a patient cohort. The clinical potential of this technique has been described, for example to assess surgical resection margins for breast (11), brain (19), and prostate (20) cancers. In lung cancer, glycerophospholipid profiling has highlighted new therapeutic approaches (21). Previously, we showed that phospholipidomics could differentiate metastatic EA from lymph node tissue (22). However, the phospholipid profile of primary EA is currently unknown, and represents a significant challenge to MSI as the complexity of this microenvironment is extraordinary and neighboring nonmalignant tissues can be phenotypically similar.

Here, we used DESI-MSI to compare the normal squamous and EA glycerophospholipid profile, using paired samples from surgical specimens. Next, we assessed the evolution of the EA glycerophospholipid profile, by comparing normal, inflamed, metaplastic, dysplastic, and neoplastic cell types, which also served as an external validation cohort. We then investigated the mechanistic basis for these glycerophospholipid signatures, by describing the corresponding fatty acid pool and genetic framework. Finally, we linked de novo lipogenesis to EA glycerophospholipid characteristics, by silencing a deregulated lipogenic gene and assessing the phospholipidomic changes in vitro.

Patients

Approval for the study was obtained via National Research Ethics Service (Ref: 04/Q0403/119) and Imperial College Healthcare Tissue Bank Institutional Review Boards (Project R14121), adhering to Declaration of Helsinki principles. Written informed consent was taken from every participant prior to collecting samples. Two independent cohorts were used, and the REMARK guidelines were followed for the recruitment of patients in this observational study (23). For cohort 1, biopsies were collected from EA surgical specimens of consecutive patients undergoing treatment of curative intent at a single tertiary referral center (St Mary's Hospital, Imperial College NHS Trust, London, United Kingdom) between April 2012 and August 2014. Samples were taken from macroscopically identified tumor and matched healthy esophageal epithelium (MHEE) 5 cm away from the tumor. Squamous tissue was selected as the initial control given the intention to apply DESI-MSI for screening surgical resection margins for cancer. Also, the selection of control tissue and distance from the tumor followed our previous work in low-molecular weight metabolites, which demonstrated field effects near esophageal cancers (24). For cohort 2, lower esophageal mucosal biopsies were accrued from patients undergoing diagnostic/therapeutic endoscopy suite at St. Mary's Hospital. Tissue types were classified into healthy esophageal epithelium (HEE, i.e., from healthy volunteers rather than cancer-adjacent normal tissue), inflamed esophageal epithelium (IEE), Barrett metaplasia (BM), Barrett dysplasia (BD), or EA prior to any treatment. All samples were collected from within 5 cm of the gastro-esophageal junction, apart from HEE, which was collected from between 15 and 5 cm. These groups were selected to assess the stepwise differences in glycerophospholipid profiles along the carcinogenic sequence of EA, and also to externally validate the findings of cohort 1. The histology of adjacent samples was independently verified by two histopathologists (R. Goldin and G. Petts) with a specialist interest in upper gastro-intestinal cancer.

Exclusion criteria for both cohorts included patients with esophageal squamous cell carcinoma, malignancy associated with any other site in the body, other gastrointestinal tract pathology, liver disease, and patients with signs/symptoms of acute infection. Demographic and clinical data including past medical history, drug history, chemotherapeutics, smoking status, and alcohol intake history were recorded for all patients (see Supplementary Tables S1 and S2). Histopathologic variables including tumor type/differentiation/grade and stage, presence of perineural/lymphovascular invasion, and lymph nodal metastases were recorded for patients in cohort 1 (see Supplementary Table S1).

Desorption electrospray MSI

Sample processing and acquisition of mass spectra

A schematic of the DESI-MSI workflow is given in Fig. 1A. A fine stream of an ionized solvent (typically methanol and water) was “sprayed” onto a cryosectioned tissue specimen, which subsequently liberates ionized target species from the sample (often lipids). These were aspirated into a mass spectrometer, which generates a mass spectrum for this particular location on the section (“pixel”). An automatic stage controller then moves to the next pixel, and a mass spectrometry “image” is built up. Tissue destruction was minimal, so the post-DESI-MSI section can be histologically stained, coregistered with the mass spectrometry image, and used to supervise discriminative analysis. Complex multivariate statistics were then needed to process the large datasets that result from the highly resolved mass spectra, from a large number of pixels, across several sampling regions, across several patients, to build up a metabolomic profile.

Figure 1.

DESI-MSI workflow and representative section. A, Schematic of the DESI-MSI workflow. B, Post-DESI-MSI hematoxylin and eosin stain of a representative esophageal biopsy section. Zoning for different tissue types demonstrated: esophageal smooth muscle (muscle, pink); Barrett metaplasia (BM; green); and Barrett dysplasia (BD; red). Scale bar, 100 μm. C, Aligned false color image of spatially resolved total ion chromatogram. D, Corresponding PCA for C. E, False color image of PI 34:1 concentration across the tissue specimen (extracted ion chromatogram, m/z 835.3542). F, Concentrations of PI 34:1 in the tissue types.

Figure 1.

DESI-MSI workflow and representative section. A, Schematic of the DESI-MSI workflow. B, Post-DESI-MSI hematoxylin and eosin stain of a representative esophageal biopsy section. Zoning for different tissue types demonstrated: esophageal smooth muscle (muscle, pink); Barrett metaplasia (BM; green); and Barrett dysplasia (BD; red). Scale bar, 100 μm. C, Aligned false color image of spatially resolved total ion chromatogram. D, Corresponding PCA for C. E, False color image of PI 34:1 concentration across the tissue specimen (extracted ion chromatogram, m/z 835.3542). F, Concentrations of PI 34:1 in the tissue types.

Close modal

All clinical tissue samples were snap-frozen in liquid nitrogen immediately upon disconnection. Samples were stored at −80°C for no longer than 1 month prior to cryosectioning at 15 μm and mounting on glass slides. All specimen processing was performed below −30°C. Clinical sample analysis was performed in random order, constant instrument/environmental settings, and in a brief time frame to avoid any batch effect influencing the results. DESI-MSI analysis was performed using an Exactive Fourier-Transform Orbitrap Mass Spectrometer (Thermo Fisher Scientific Inc.) controlled by XCalibur 2.1 software. Mass spectrometry data were acquired in negative ion mode in the m/z150–1,000 range. The spatial resolution for the imaging experiments was set to 75 μm, corresponding to X/Y dimensions of the image pixels. Full protocol for optimization, precision measurements, and spectra acquisition are reported in a previous study (25).

Tissue sections were poststained for histologic confirmation and direct comparison was made with the mass spectrometry image for cell/tissue-specific data extraction. All samples were stained with hematoxylin and eosin prior to digital imaging with a high-resolution digital microscope (NanoZoomer 2.0-HT digital slide scanner) and assessed by two blinded, consultant histopathologists for histologic mapping of cell/tissue types within the specimens.

Tissue-specific mass spectra extraction

Raw mass spectrometric data were converted to imzML format via imzML Converter (version 1.0.5; ref. 26) and imported into MATLAB (R2014a) for feature extraction, preprocessing, and data analysis using an in-house bioinformatics platform (27). Each MSI image was composed of multiple pixels representing individual mass spectra. To extract relevant mass spectra of specific cell/tissue types the MSI image must be compared with its matched histologic image. As previously described, the histologic image was assessed by an independent consultant histopathologist for identification and spatial mapping of specific cell/tissue types, which was recorded on Digital Imaging Software (NanoZoomer, Hamamatsu).

The histologic and MSI images were then aligned using an in-house–developed automated affine image transformation (translation, rotation, and scaling) algorithmic based on a gradient descent optimization approach (27). The coregistration of the matched histology and MSI images permitted selection of mass spectra of specific tissue types by highlighting a corresponding two-dimensional area on the matched histology image. This process was repeated for each sample to extract mass spectra of each specific tissue type to populate a composite database. Within the database, the mass spectra of each tissue type was also categorized with respect to its sample of origin to perform intersample comparisons.

Data preprocessing

Negative ion mode data were acquired between the 150–1,000 m/z range. Analysis of glycerophospholipids was performed on the 600–1,000 m/z range. Analysis of fatty acids was performed on the 150–400 m/z range. Subsequent data preprocessing of the raw data was then performed on the selected m/z range of interest. Despite daily mass calibration of the exactive mass spectrometer, nonlinear mass shifts of common peaks were evident across multiple samples. To perform comparative analysis of multiple samples, these mass shifts were corrected by means of an in-house peak alignment algorithm (dynamic programming, 8ppm matched; ref. 28). To account for differential lipid density owing to differences in cell morphology, spectra were normalized to the total ion count. Therefore, the only differences were in the ratios of the spectral features (corresponding to lipids in this case) found within that mass range. The spectra were then denoised; peaks were identified as noise and eliminated if present in <50% of representative mass spectra of any tissue type. Finally, the mean average of selected pixels of each tissue type within each patient sample was calculated and stored as a single data unit.

Statistical analysis

All analyses were conducted in MATLAB, using a previously published toolbox (27). To illustrate the dataset in reduced dimensional space, principal components analysis (PCA) was initially used as the unsupervised technique. As described, each tissue type from each sample was average across several sampling regions, such that each point indicated one tissue type from one sample. The supervised approach was recursive maximum margin criterion, and internally cross-validated models were generated using k-fold/leave one out cross-validation (29). Misclassifications were visualized using confusion matrices. Glycerophospholipid/fatty acid annotations were made using LipidMAPS (30), after exclusion of isotopologues. The mean intensity values of lipids in different tissue subclasses (e.g., cancer vs. proximal healthy epithelium) were compared by log2 transformed mean fold change. Lipid intensities between tissue subclasses were compared for statistically significant differences by ANOVA, with P values adjusted to q values with a FDR of 0.001, to reduce multiplicity error. Overall class/saturation/acyl chain length comparisons of glycerophospholipid between tissues were compared using ANOVA with Tukey post hoc analysis. Further explanation of the statistical approach can be found in the Supplementary Materials and Methods.

Quantitative reverse transcriptase-PCR

The MIQE guidelines were followed. Endoscopic biopsies (2–3 each) of EA and MHEE were collected as for DESI-MSI (replicate samples of cohort 1). The samples were mounted in optimal cutting temperature (OCT) compound (ThermoLife), flank cryosectioned, and microdissected to achieve a target cellularity of >90%. The OCT was trimmed and the sample homogenized in TRIzol (ThermoLife) using a three-step protocol involving (i) a hand-held oscillating pestle, (ii) 30 seconds steel bead-beating, and (iii) passage of the homogenate through a Qiashredder (Qiagen). This optimized protocol was necessary to acquire adequate yield (>50 ng) from small squamous samples while maintaining RNA integrity (RIN > 7). RNA was then fractionated in according the TRIzol instructions, and then additionally purified using RNEasy columns, with an extra wash step with both RW1 and RPE buffers. cDNA was made using SuperScript III (ThermoLife) and quantified using PowerSybr master mix and AB7900 thermal cycler, all used according to the manufacturer's instructions. Controls without template did not amplify. GAPDH was selected as the internal control gene as this was the most stable across an initial run of 10 samples. Oligonucleotide details are provided in the Supplementary Materials and Methods.

Cell culture and transfection

The FLO1 EAC cell line was purchased from the European Collection of Authenticated Cell Cultures (Public Health England) at the start of the study, who had authenticated the line using short tandem repeat profiling. Cultures were Mycoplasma tested every month (last tested January 10, 2020), and all experiments were conducted within 10 passages of the original stock. Cultures were established in conditions recommended by the accompanying literature. For RNA silencing experiments, cells were seeded onto glass cover slips and transiently transfected with an ACLY-targeting siRNA, or nontargeting siRNA (“ControlScramble”, Silencer Select, Ambion, and control RNA 1), or water (“ControlVector”), using oligofectamine (Life Technologies) according to the manufacturer's instructions. After 48 hours, cells were washed twice in 150 mmol/L ammonium acetate, and snap frozen on the cover slip prior to direct DESI-MSI. Silencing efficiency was checked using immunoblotting of replicate wells as described previously (31), using the ACLY primary antibody 4223 (Cell Signaling Technology). Selection of siRNA was based on sequence alignment to minimize off-target effects.

IHC

The expression of the key enzymes involved in the de novo lipogenesis pathway (ACLY, ACACA, FASN, SCD, and ELOVL1) was assessed in EA and MHEE in 20 patients from cohort 1. Fresh tissue samples adjacent to those taken for DESI-MSI and qPCR were fixed in formalin and mounted in wax. After sectioning, specimens were dewaxed, hydrated, and retrieved in bond epitope retrieval solution 1 for 20 minutes or solution 2 for 40 minutes at 100°C. Peroxidases were blocked, and the specimens were incubated with optimized concentrations of antibodies for 30–120 minutes (primary antibody details are in the Supplementary Materials and Methods). Primary antibody binding was visualized using Bond Polymer Refine Detection (DS9800, Leica). All staining procedures were carried out on a Leica Bond Autostainer.

Immunoreactivity scoring was independently undertaken by two histopathologists (R. Goldin and G. Petts). The analysis was confined to cancer cells in the tumor samples versus squamous cells in the normal esophageal mucosa. The H-score method was used: score 0, 1, 2, and 3 for intensity (negative, weak, moderate, and strong), multiplied by the percentage of staining cells, out of a maximum of 300. Median scores were calculated for the cancer and healthy mucosal groups for each lipogenic enzyme and compared by Mann–Whitney U test with P < 0.05 to define statistical significance.

The glycerophospholipid signature of esophageal adenocarcinoma

DESI-MSI was able to distinguish the tissue types within the complex EA microenvironment (see Fig. 1B). A representative analysis is provided in Fig. 1BF, demonstrating differentiation between Barrett metaplasia, Barrett dysplasia, and smooth muscle within a single section. Tissue architecture was preserved (Fig. 1B). A multivariate model derived from across the glycerophospholipid mass range (m/z 600–1,000) segregated these groups in reduced dimensional space (Fig. 1C and D). Specific discriminating lipids could be quantitatively visualized. For example, PI 34:1 was greatly enriched in Barrett cells compared with adjacent smooth muscle, and most greatly in dysplastic Barrett (m/z 835.5342; P = 3.1 × 10−32; see Fig. 1E and F).

The glycerophospholipid signature of esophageal adenocarcinoma was characterized using two cohorts. Initially (cohort 1), paired surgical resection biopsies (EA and MHEE) from 33 consecutive patients were profiled for glycerophospholipid differences using DESI-MSI. This sampling approach was selected as the intended application was for operative margin analysis, and to overcome regional metabolomic differences in the tumor by sampling a larger specimen. Next, a second cohort was sampled using endoscopic biopsies. This allowed specimens to be acquired from healthy volunteers and patients with Barrett metaplasia/dysplasia. It also allowed us to validate findings from the first cohort, investigate the performance of the technique on small specimens, and demonstrate the second clinical application of facilitated diagnosis in the endoscopy suite.

Cohort 1: discovery

The analytic pipeline calculated average mass spectra for each patient, by synthesizing data within and between histologically verified sampling zones. Typically, for each specimen, 10–15 zones with approximately 100–200 pixels each were sampled (one pixel = one mass spectra), such that each phospholipidome represented at least 10,000 mass spectra. The demographics and case characteristics of this cohort of patients are included in Supplementary Table S1.

Unsupervised multivariate analysis of mass spectra in the m/z 600–1,000 range demonstrated clear segregation between EA and MHEE (Fig. 2A). Separation was apparent in the first component of the PCA, which explained 22.5% of the variance. This separation was more pronounced on supervised analysis (Fig. 2B). After internal cross-validation, the model correctly classified 93.9% of EA and MHEE (see Fig. 2C; Supplementary Fig. S1A), with two EA samples were incorrectly classified. The AUROC was 0.97 (see Supplementary Fig. S1B).

Figure 2.

Glycerophospholipid signatures of esophageal adenocarcinoma and MHEE (cohort 1). A, PCA of averaged mass spectra (m/z 600–1,000) of esophageal adenocarcinoma (EA; red) versus MHEE (green) of patients in cohort 1 (each dot, averaged spectra from one tissue type from one patient; each patient provided both EA and MHEE samples). B, RMMC supervised analysis score plot. C, Leave one out cross-validated RMMC score plot. Relative abundance of glycerophospholipids (GPL) by class (D), acyl chain length (E), and desaturations (F). Glycerophospholipid species that were significantly different (q < 0.001) by class (G), acyl chain length (H), and desaturations (I). Groups were compared by ANOVA (with Tukey HSD): *, P < 0.05; **, P < 0.01; ***, P < 0.001. PA, phosphatidic acid; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserines.

Figure 2.

Glycerophospholipid signatures of esophageal adenocarcinoma and MHEE (cohort 1). A, PCA of averaged mass spectra (m/z 600–1,000) of esophageal adenocarcinoma (EA; red) versus MHEE (green) of patients in cohort 1 (each dot, averaged spectra from one tissue type from one patient; each patient provided both EA and MHEE samples). B, RMMC supervised analysis score plot. C, Leave one out cross-validated RMMC score plot. Relative abundance of glycerophospholipids (GPL) by class (D), acyl chain length (E), and desaturations (F). Glycerophospholipid species that were significantly different (q < 0.001) by class (G), acyl chain length (H), and desaturations (I). Groups were compared by ANOVA (with Tukey HSD): *, P < 0.05; **, P < 0.01; ***, P < 0.001. PA, phosphatidic acid; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserines.

Close modal

A total of 192 glycerophospholipids were identified, with disparate abundances throughout the mass range (see Supplementary Fig. S1C). In MHEE, phosphatidylethanolamines and phosphatidylinositols were the most abundant groups of glycerophospholipids, followed by phosphatidylserines, phosphatidylglycerols, and phosphatidic acids (see Fig. 2D). In EA there were significantly more phosphatidylglycerols (ANOVA with Tukey, P < 0.0001) and less phosphatidic acids (P < 0.05) and phosphatidylethanolamines (P < 0.001), compared with MHEE (see Fig. 2D).

The predominant glycerophospholipid chain lengths had even numbers of carbon (34, 36, 38, and 40, i.e., various combinations of the C16:x, C18:x, and C20:x fatty acids, see Fig. 2E). Compared with MHEE, there was a tendency toward longer acyl chain lengths with significantly more EA species with 37 (P < 0.05), 39 (P < 0.01), 40 (P < 0.0001), 42 (P < 0.01), and 43 (P < 0.0001) total acyl carbons, and significantly less 33 (P < 0.001) and 34 (P < 0.0001) total acyl carbons (see Fig. 2E). Significant differences also existed in odd chain glycerophospholipids despite very low abundances. Across the tissue types, glycerophospholipid most frequently had 1, 2, or 4 desaturations (see Fig. 2F). There was significantly less saturated and monounsaturated acyls in EA compared with MHEE (P < 0.01), and significantly more polyunsaturated acyls (P < 0.01; see Fig. 2F). The glycerophospholipid species that most significantly contributed to phosphatidylethanolamine/phosphatidylglycerol, total acyl length, and total acyl desaturation class differences are provided in Fig. 2GI, respectively.

Specific differences in glycerophospholipid species are provided in Supplementary Table S3, together with significantly different plasmalogens (i.e., monoacylglycerols) that were identified. Of the 20 most significantly enriched glycerophospholipid species (ANOVA with FDR), 10 were phosphatidylglycerols, and nine were enriched in EA by at least 2 log2 (fold change). Additional validation of glycerophospholipid annotation by MS2 (using the Exactive Orbitrap FTMS) was also possible for abundant molecular ions and is also provided in Supplementary Table S3. Examples of molecular and product ion scans are provided in Supplementary Fig. S2A–S2H.

Cohort 2: validation and stepwise differences

To validate these findings and understand whether stepwise glycerophospholipid differences occur with different stages of carcinogenesis (see Fig. 3A), endoscopic biopsies from relevant patients were profiled using DESI-MSI with histopathologic confirmation (HEE = 33; IEE = 8; BM = 26; BD = 7; EA = 10). Supplementary Table S2 demonstrates the case characteristics of these patients; demographics were generally matched, except that the IEE cases were younger and acid-suppression medication was higher in the BM group.

Figure 3.

Glycerophospholipid (GPL) signatures of esophageal and various control tissues (cohort 2). A, Representative histology of HEE, IEE, Barrett metaplasia, Barrett dysplasia, and esophageal adenocarcinoma. Color coding for rest of figure is provided. B, PCA of averaged mass spectra (m/z 600–1,000) of the five tissue types of patients in cohort 2 (each dot, averaged spectra from one tissue type from one patient; each patient provided one tissue type only). C, RMMC supervised analysis score plot. D, Leave one out cross-validated RMMC score plot as per confusion matrix. Relative abundances of glycerophospholipids (m/z 600–1,000 range) grouped in terms of class (E), acyl chain length (F), and desaturations (G). Groups were compared by ANOVA (with Tukey HSD): *, P < 0.05; **, P < 0.01; ***, P < 0.001. PA, phosphatidic acid; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserines.

Figure 3.

Glycerophospholipid (GPL) signatures of esophageal and various control tissues (cohort 2). A, Representative histology of HEE, IEE, Barrett metaplasia, Barrett dysplasia, and esophageal adenocarcinoma. Color coding for rest of figure is provided. B, PCA of averaged mass spectra (m/z 600–1,000) of the five tissue types of patients in cohort 2 (each dot, averaged spectra from one tissue type from one patient; each patient provided one tissue type only). C, RMMC supervised analysis score plot. D, Leave one out cross-validated RMMC score plot as per confusion matrix. Relative abundances of glycerophospholipids (m/z 600–1,000 range) grouped in terms of class (E), acyl chain length (F), and desaturations (G). Groups were compared by ANOVA (with Tukey HSD): *, P < 0.05; **, P < 0.01; ***, P < 0.001. PA, phosphatidic acid; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserines.

Close modal

To validate the findings from cohort 1, the phospholipidomes of HEE and EA only were initially compared (see Supplementary Fig. S3). Unsupervised multivariate analysis of mass spectra in the m/z 600–1,000 range demonstrated separation of the data points along the first principal component (Supplementary Fig. S3A), with clear differences throughout the mass range (Supplementary Fig. S3B). The derived cross-validated recursive maximum margin criterion (RMMC) model from these data correctly identified 100% of both tissue types (Supplementary Fig. S3C), providing an AUROC of 1 (Supplementary Fig. S3D). The leading difference was an abundance of phosphatidylglycerols, with longer acyl chain and more desaturations. These data verify MSI-based phospholipidomics as an accurate means of esophageal tissue recognition.

To assess for stepwise differences through transformation, phospholipidomes of the remaining tissue types were added to the analysis. Unsupervised multivariate analysis of mass spectra in the m/z 600–1,000 range also demonstrated separation of data points along the first principal component between HEE/IEE and BM/BD/EA (Fig. 3B), suggesting the leading glycerophospholipid profiles' differences are due to the squamous or columnar phenotypes. Supervised analysis using recursive maximum margin criterion verified segregation based on tissue-of-origin (Fig. 3C). However, there was further clustering of EA from BM/BD along the second component, associating further glycerophospholipid reprogramming with transformation. Internal cross-validation incorrectly classified one squamous sample as columnar; however, BD was more frequently misclassified as either BM or EA (Fig. 3D).

The relative abundance of the 196 identified glycerophospholipids (m/z 600–1,000 range) were compared, in terms of glycerophospholipid class, acyl chain length, and desaturations (Fig. 3EG). The relative quantity of phosphatidylglycerols was higher in EA/BD/BM compared with HEE/IEE (P < 0.01). The relative quantity of phosphatidylinositols was higher in BM/BD compared with HEE (P < 0.05). The relatively quantity of phosphatidylserines was lower in BM/BD/EA compared with HEE (P < 0.0001) and IEE (P < 0.05).

Overall, acyl chain length was longer in BM/BD/EA, with significantly greater concentrations of 37, 38, and 40 carbon lengths, and less chains of 34 (Fig. 3F). Acyl chains from BM/BD/EA had more desaturations, including significantly more having four or more double bonds, and significantly less having less than two (Fig. 3G). In summary, glycerophospholipids of BM/BD/EA were enriched for phosphatidylglycerols, and had longer acyl chains with more desaturations.

To understand whether the glycerophospholipid signature changes during progression to invasive cancer, we compared univariate phosphatidylglycerol characteristics between EA with BM, as this class was most enriched in cancer (Fig. 3E; Supplementary Table S4). Comparison of individual glycerophospholipids showed a significant increase in six long-chain polyunsaturated phosphatidylglycerols (q < 0.001), with phosphatidylglycerol (42:8) having the greatest increase (log2 fold change 6.1; ANOVA q = 8.6 × 10−4). The other enriched phosphatidylglycerols were phosphatidylglycerol (38:4), (38:5), (38:6), (40:5), and (40:6).

Drivers of the EA glycerophospholipid signature

The EA fatty acid profile

Long chain fatty acids of 12–26 carbon atoms are used to make glycerophospholipid acyl chains in humans (8), so profiling the fatty acid pool may offer mechanistic insights into glycerophospholipid acyl characteristics. The relative abundance of fatty acids found within the m/z150–400 range were compared between EA and MHEE from cohort 1. In unsupervised analysis using data from this mass range only, there was separation of the data points along the third component, which explained 8.9% of the variance (Fig. 4A). This separation was more pronounced in supervised analysis (Fig. 4B). Internal cross-validation correctly classified 87.9% of EA and 90.9% of MHEE based on the fatty acid profiles of the samples, with an AUC of 0.96 (Fig. 4C). Fatty acid acyls had significantly less 16-carbon chains in EA, and significantly more with longer chains (see Fig. 4D and E). There were also significantly less saturated fatty acid acyls in EA and significantly more desaturated acyls (see Fig. 4F and G).

Figure 4.

Fatty acid signatures of esophageal adenocarcinoma and MHEE in cohort 1. A, PCA of averaged mass spectra (m/z 150–400) of esophageal adenocarcinoma (EA; red) versus MHEE (green) of patients in cohort 1 (each dot, averaged spectra from one tissue type from one patient; each patient provided both EA and MHEE samples). B, RMMC supervised analysis score plot. C, Leave one out cross-validated RMMC score plot. Relative abundance of fatty acids (m/z 150–400) grouped by acyl chain length (D) and desaturations (F). Fatty acid (FA) species that were significantly different (q < 0.001) by acyl chain length (E) and desaturations (G). Groups were compared by ANOVA (with Tukey HSD). *, P < 0.05; ***, P < 0.001.

Figure 4.

Fatty acid signatures of esophageal adenocarcinoma and MHEE in cohort 1. A, PCA of averaged mass spectra (m/z 150–400) of esophageal adenocarcinoma (EA; red) versus MHEE (green) of patients in cohort 1 (each dot, averaged spectra from one tissue type from one patient; each patient provided both EA and MHEE samples). B, RMMC supervised analysis score plot. C, Leave one out cross-validated RMMC score plot. Relative abundance of fatty acids (m/z 150–400) grouped by acyl chain length (D) and desaturations (F). Fatty acid (FA) species that were significantly different (q < 0.001) by acyl chain length (E) and desaturations (G). Groups were compared by ANOVA (with Tukey HSD). *, P < 0.05; ***, P < 0.001.

Close modal

Genetic basis for glycerophospholipid signature and contribution of de novo lipogenesis

We hypothesized that glycerophospholipid class-switch was driven by corresponding changes in gene expression. To test this hypothesis and generate candidates, we checked whether the Kyoto Encyclopedia of Genes and Genomes glycerophospholipid gene set was significantly altered in an archived transcriptomic dataset for EA and HEE (GSE 26886). Overall, EA was significantly enriched for glycerophospholipid synthetic genes (see Fig. 5A). To explore this further, we extracted mRNA from endoscopic biopsies of HEE and EA (n = 20 each), and quantified the expression of the most significantly altered candidates (see Fig. 5B and C). Genes involved in phosphatidylglycerol synthesis, both by diacylglycerol phosphorylation and by lyso-phosphatidylglycerol acylation, were strongly enriched in EA compared with HEE [median fold change 2.5 (LPGAT) and 8.3 (PGS1) both P < 0.001]. The final synthetic step in phosphatidylglycerol metabolism, dimerization by CRLS1 to cardiolipin, was also upregulated in EA.

Figure 5.

The genetic framework of esophageal adenocarcinoma glycerophospholipid metabolism. A, Gene set enrichment score plot of the glycerophospholipid metabolism gene set in GSE 26886. B, Overview of glycerophospholipid metabolism with schematics of the various glycerophospholipid classes. C, Relative glycerophospholipid gene expression between esophageal adenocarcinoma (cancer, EA) and MHEE (normal). D, Representative IHC sections of the five fatty acid metabolism genes in EA and MHEE. Scale bar applies to all images and represents 100 μm. E, Relative abundance of acyl chain lengths in FLO1 EA cells transfected with siRNA targeting ACLY and relevant controls. P values calculated with Mann–Whitney U test for qPCR experiments and ANOVA with Tukey HSD for glycerophospholipid profiling experiments. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.001. CDP-DAG cytidine diphosphate diacylglycerol; DAG, diacylglycerol; DAG-3P, diacylglycerol-triphosphate; PA, phosphatidic acid, PC, phosphatidylcholine; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserines.

Figure 5.

The genetic framework of esophageal adenocarcinoma glycerophospholipid metabolism. A, Gene set enrichment score plot of the glycerophospholipid metabolism gene set in GSE 26886. B, Overview of glycerophospholipid metabolism with schematics of the various glycerophospholipid classes. C, Relative glycerophospholipid gene expression between esophageal adenocarcinoma (cancer, EA) and MHEE (normal). D, Representative IHC sections of the five fatty acid metabolism genes in EA and MHEE. Scale bar applies to all images and represents 100 μm. E, Relative abundance of acyl chain lengths in FLO1 EA cells transfected with siRNA targeting ACLY and relevant controls. P values calculated with Mann–Whitney U test for qPCR experiments and ANOVA with Tukey HSD for glycerophospholipid profiling experiments. *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.001. CDP-DAG cytidine diphosphate diacylglycerol; DAG, diacylglycerol; DAG-3P, diacylglycerol-triphosphate; PA, phosphatidic acid, PC, phosphatidylcholine; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserines.

Close modal

We then checked whether genes involved in fatty acid synthesis were also deregulated in EA, using IHC. This revealed robust EA expression of ACYL, FASN, EVOVL1, ACACA, and SCD that was at least equivalent to or greater than MHEE (see Fig. 5D; Supplementary Fig. S4A). Given that both the fatty acid and glycerophospholipid-acyl pools of EA had similar characteristics, and that genes involved in de novo lipogenesis were active in EA, it was hypothesized that de novo lipogenesis contributes to glycerophospholipid acyl reprogramming. To test this, ACYL was silenced in FLO1 EA cells in vitro, and the resulting glycerophospholipid signature was measured with DESI-MSI (Supplementary Fig. S4B and S4C). ACYL channels carbon from the citrate cycle to acetyl CoA, thus providing the materials for acyl elongation, and is considered the first committed step in carbon shuttling to lipid anabolism (32). After 72 hours, there were significantly less glycerophospholipids with total acyl chain length of 40, and significantly more with 36 carbons (see Fig. 5E). In summary, genes involved in glycerophospholipid and de novo fatty acid synthesis were broadly enriched in EA, and impairing de novo lipogenesis in EA cells reverted glycerophospholipid acyls toward a normal phenotype.

In this study, we combined DESI-MSI lipidomic profiling with gene expression and perturbation studies to describe the glycerophospholipid signature of EA, and its genetic basis. In two consecutive tissue series, sampled surgically and then endoscopically, multivariate models–based derived from glycerophospholipid profiles were highly discriminant for EA compared with squamous and other control tissue (AUROC = 0.97 and 1). These results suggest DESI-MSI can differentiate tissue types in the malignant esophagus. Potential clinical applications include objective diagnosis and intraoperative margin assessment, especially given recent reports of rapid processing times (33).

There was stepwise enrichment of phosphatidylglycerols from normal to malignant tissue samples. Small concentrations of phosphatidylglycerols can have potent effects on signaling (14, 34, 35), and phosphatidylglycerols are usually at trace concentrations in mammalian tissues. In EA, phosphatidylglycerols were the third most abundant glycerophospholipid class, and how this significant increase in phosphatidylglycerol affects EA signaling requires further study. Phosphatidylglycerols are also the precursors of cardiolipins, which contribute to mitochondrial functionality and are frequently altered in cancer states (36–38). Phosphatidylinositols were particularly enriched in BM/BD. Given the interest in PI3K signaling to EA (39), this observation also warrants further investigation. The decrease in phosphatidylserine concentration may also be protumorigenic, as phosphatidylserine is a proapoptotic signaling molecule and a chemoattractant for macrophages and other immune cells (40). The significance of the slight decrease in phosphatidic acids is less clear and may reflect increased lipase activity, leading to phosphatidylglycerol and phosphatidylinositol enrichment.

Acyl chains of EA glycerophospholipids had more desaturations and longer lengths, which probably reflect corresponding changes in the fatty acid pool. Genes that de novo synthesize, extend, and desaturate fatty acids were strongly expressed in EA. In ovarian cancer, desaturases support cancer stemness and drive nf-κB signaling (41). Elongase activity and longer acyl chain lengths have been associated with a procancer phenotype in lung (42, 43), prostate (44), and breast cancers (45), implying that these characteristics confer ligand activity. Recently it was demonstrated that different phosphatidylglycerol species have contrasting effects on mouse skin keratinocyte proliferation (14). Similar species-specific effect has been demonstrated for cardiolipins (38). In this dataset some EA lipids were up to 70× enriched, and an important next step will be the functional annotation of these discriminatory lipids within EA signaling.

A strength of this article was to begin to describe the underlying genetic and mechanistic basis of phospholipid reprogramming. Genes that control de novo lipogenesis and glycerophospholipid synthesis were generally upregulated in EA, implying diversion to the final products of glycerophospholipid metabolism (e.g., phosphatidylinositols and cardiolipins). Our finding that ACLY silencing reduces glycerophospholipid chain length supports the hypothesis that these phenotypes are partly explained by acetyl CoA–derived de novo lipogenesis and/or elongation (32), which is also supported by the observed robust expression of relevant lipogenic genes. Recently, mTORC2 was shown to be a master regulator of glycerophospholipid synthesis, and is hyperactive in EA (46); however, the complete coordination of de novo lipogenesis is likely multifactorial. In addition, the promoter of PTDSS1 is frequently mutated in EA (13), which suggests lipid reprogramming is selected in oncogenesis. It should be stressed that the ACLY experiment constitutes a technical and biological proof-of-principle, and a more comprehensive functional and mechanistic assessment of the effects of deregulated lipogenic genes is needed.

Additional strengths include the use of complementary independent cohorts of patient samples to corroborate our lipidomic findings; transparent details of patient selection; the use of multiple control samples that where either matched to the tumor from the resection specimen (cohort 1) or matched in terms of demographics, medication, and lifestyle factors (cohort 2); gold standard determination by two expert histopathologists; and a robust methodology for data analysis and interpretation. This study's limitations include the sample sizes of BM/BD and the single analytic platform for lipidomic profiling. In addition, this study did not assess the signatures of minor lipid classes such as cholesterols, ceramides, and sphingomyelins. Future studies should use externally calibrated tandem mass spectrometry validation, supported by molecular studies to test the relevance of the lipid species and genes on the overall phenotype. Specific inhibitors are available for lipogenic genes (47, 48), and thus these findings may indicate new therapeutic avenues.

In conclusion, DESI-MSI can objectively recognize adenocarcinoma in the esophagus. The EA phospholipidome is greatly enriched for long-chain, polyunsaturated phosphatidylglycerols, and genetic studies suggest an orchestrated mechanism linked to de novo lipogenesis. The functional effect of the discriminatory lipids remains to be determined.

K. Veselkov is a consultant (paid consultant) at Waters Corp. Z. Takáts is a consultant (paid consultant) at and reports receiving a commercial research grant from Waters Corporation. No potential conflicts of interest were disclosed by the other authors.

Conception and design: N. Abbassi-Ghadi, S.S. Antonowicz, S. Kumar, K. Veselkov, Z. Takáts, G.B. Hanna

Development of methodology: N. Abbassi-Ghadi, S.S. Antonowicz, N. Strittmatter, K. Veselkov, Z. Takáts, G.B. Hanna

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): N. Abbassi-Ghadi, S.S. Antonowicz, S. Kumar, J. Huang, E.A. Jones, N. Strittmatter, H. Kudo, S. Court

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): N. Abbassi-Ghadi, S.S. Antonowicz, J.S. McKenzie, S. Kumar, E.A. Jones, G. Petts, K. Veselkov, R. Goldin, Z. Takáts, G.B. Hanna

Writing, review, and/or revision of the manuscript: N. Abbassi-Ghadi, S.S. Antonowicz, J.S. McKenzie, S. Kumar, J.M. Hoare, R. Goldin, Z. Takáts, G.B. Hanna

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): N. Abbassi-Ghadi, S. Kumar, H. Kudo, J.M. Hoare

Study supervision: K. Veselkov, Z. Takáts, G.B. Hanna

Other (collection of specimens): S. Court

Other (sample collection and case identification): J.M. Hoare

Other (supervision and design of machine learning and informatics strategies for improved information recovery from MSI data): K. Veselkov

Other (the histology and IHC analysis): R. Goldin

The authors would like to acknowledge funding from the European Research Council (DESI-JeDI Imaging Starting Grant; MASSLIP Consolidator Grant), the National Institute of Health Research (Imperial Biomedical Research Centre), and the National Institute of Health Research–London In Vitro Diagnostics.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Torre
LA
,
Bray
F
,
Siegel
RL
,
Ferlay
J
,
Lortet-Tieulent
J
,
Jemal
A
. 
Global cancer statistics, 2012
.
CA Cancer J Clin
2015
;
65
:
87
108
.
2.
Enzinger
PC
,
Mayer
RJ
. 
Esophageal cancer
.
N Engl J Med
2003
;
349
:
2241
52
.
3.
Chadwick
G
,
Varagunam
M
,
Maynard
N
,
Brand
C
,
Cromwell
D
,
Riley
S
, et al
National oesophago-gastric cancer audit 2016
.
Accessed May 12, 2020 at
www.nogca.org.uk.
4.
Hanahan
D
,
Weinberg
RA
. 
The hallmarks of cancer
.
Cell
2000
;
100
:
57
70
.
5.
Kroemer
G
,
Pouyssegur
J
. 
Tumor cell metabolism: cancer's Achilles' heel
.
Cancer Cell
2008
;
13
:
472
82
.
6.
Holmes
E
,
Wilson
ID
,
Nicholson
JK
. 
Metabolic phenotyping in health and disease
.
Cell
2008
;
134
:
714
7
.
7.
Wishart
DS
,
Knox
C
,
Guo
AC
,
Eisner
R
,
Young
N
,
Gautam
B
, et al
HMDB: a knowledgebase for the human metabolome
.
Nucleic Acids Res
2009
;
37
:
D603
10
.
8.
Wenk
MR
. 
The emerging field of lipidomics
.
Nat Rev Drug Discov
2005
;
4
:
594
610
.
9.
Rahman
SMJ
,
Gonzalez
AL
,
Li
M
,
Seeley
EH
,
Zimmerman
LJ
,
Zhang
XJ
, et al
Lung cancer diagnosis from proteomic analysis of preinvasive lesions
.
Cancer Res
2011
;
71
:
3009
17
.
10.
Eberlin
LS
,
Norton
I
,
Dill
AL
,
Golby
AJ
,
Ligon
KL
,
Santagata
S
, et al
Classifying human brain tumors by lipid imaging with mass spectrometry
.
Cancer Res
2011
;
72
:
645
54
.
11.
Guenther
S
,
Muirhead
LJ
,
Speller
AVM
,
Golf
O
,
Strittmatter
N
,
Ramakrishnan
R
, et al
Spatially resolved metabolic phenotyping of breast cancer by desorption electrospray ionization mass spectrometry
.
Cancer Res
2015
;
75
:
1828
37
.
12.
Dória
ML
,
McKenzie
JS
,
Mroz
A
,
Phelps
DL
,
Speller
A
,
Rosini
F
, et al
Epithelial ovarian carcinoma diagnosis by desorption electrospray ionization mass spectrometry imaging
.
Sci Rep
2016
;
6
:
39219
.
13.
Frankell
AM
,
Jammula
S
,
Li
X
,
Contino
G
,
Killcoyne
S
,
Abbas
S
, et al
The landscape of selection in 551 esophageal adenocarcinomas defines genomic biomarkers for the clinic
.
Nat Genet
2019
;
51
:
506
16
.
14.
Xie
D
,
Seremwe
M
,
Edwards
JG
,
Podolsky
R
,
Bollag
WB
. 
Distinct effects of different phosphatidylglycerol species on mouse keratinocyte proliferation
.
PLoS One
2014
;
9
:
e107119
.
15.
Sommer
U
,
Herscovitz
H
,
Welty
FK
,
Costello
CE
. 
LC-MS-based method for the qualitative and quantitative analysis of complex lipid mixtures
.
J Lipid Res
2006
;
47
:
804
14
.
16.
Pulfer
M
,
Murphy
RC
. 
Electrospray mass spectrometry of phospholipids
.
Mass Spectrom Rev
2003
;
22
:
332
64
.
17.
McDonnell
LA
,
Heeren
RMA
. 
Imaging mass spectrometry
.
Mass Spectrom Rev
2007
;
26
:
606
43
.
18.
Gode
D
,
Volmer
DA
. 
Lipid imaging by mass spectrometry–a review
.
Analyst
2013
;
138
:
1289
.
19.
Jarmusch
AK
,
Pirro
V
,
Baird
Z
,
Hattab
EM
,
Cohen-Gadol
AA
,
Cooks
RG
. 
Lipid and metabolite profiles of human brain tumors by desorption electrospray ionization-MS
.
Proc Natl Acad Sci U S A
2016
;
113
:
1486
91
.
20.
Banerjee
S
,
Zare
RN
,
Tibshirani
RJ
,
Kunder
CA
,
Nolley
R
,
Fan
R
, et al
Diagnosis of prostate cancer by desorption electrospray ionization mass spectrometric imaging of small metabolites and lipids
.
Proc Natl Acad Sci U S A
2017
;
114
:
3334
9
.
21.
Gouw
AM
,
Eberlin
LS
,
Margulis
K
,
Sullivan
DK
,
Toal
GG
,
Tong
L
, et al
Oncogene KRAS activates fatty acid synthase, resulting in specific ERK and lipid signatures associated with lung adenocarcinoma
.
Proc Natl Acad Sci U S A
2017
;
114
:
4300
5
.
22.
Abbassi-Ghadi
N
,
Golf
O
,
Kumar
S
,
Antonowicz
S
,
McKenzie
JS
,
Huang
J
, et al
Imaging of esophageal lymph node metastases by desorption electrospray ionization mass spectrometry
.
Cancer Res
2016
;
76
:
5647
56
.
23.
McShane
LM
,
Altman
DG
,
Sauerbrei
W
,
Taube
SE
,
Gion
M
,
Clark
GM
. 
Reporting recommendations for tumor marker prognostic studies (REMARK)
.
J Natl Cancer Inst
2005
;
97
:
1180
4
.
24.
Yakoub
D
,
Keun
HC
,
Goldin
R
,
Hanna
GB
. 
Metabolic profiling detects field effects in nondysplastic tissue from esophageal cancer patients
.
Cancer Res
2010
;
70
:
9129
36
.
25.
Abbassi-Ghadi
N
,
Jones
EA
,
Veselkov
KA
,
Huang
J
,
Kumar
S
,
Strittmatter
N
, et al
Repeatability and reproducibility of desorption electrospray ionization-mass spectrometry (DESI-MS) for the imaging analysis of human cancer tissue: a gateway for clinical applications
.
Anal Methods
2015
;
7
:
71
80
.
26.
Schramm
T
,
Hester
Z
,
Klinkert
I
,
Both
J-P
,
Heeren
RMA
,
Brunelle
A
, et al
imzML—a common data format for the flexible exchange and processing of mass spectrometry imaging data
.
J Proteomics
2012
;
75
:
5106
10
.
27.
Veselkov
KA
,
Mirnezami
R
,
Strittmatter
N
,
Goldin
RD
,
Kinross
J
,
Speller
AVM
, et al
Chemo-informatic strategy for imaging mass spectrometry-based hyperspectral profiling of lipid signatures in colorectal cancer
.
Proc Natl Acad Sci U S A
2014
;
111
:
1216
21
.
28.
Inglese
P
,
McKenzie
JS
,
Mroz
A
,
Kinross
J
,
Veselkov
K
,
Holmes
E
, et al
Deep learning and 3D-DESI imaging reveal the hidden metabolic heterogeneity of cancer
.
Chem Sci
2017
;
8
:
3500
11
.
29.
Zhang
H
,
Yang
S
,
Guo
L
,
Zhao
Y
,
Shao
F
,
Chen
F
. 
Comparisons of isomiR patterns and classification performance using the rank-based MANOVA and 10-fold cross-validation
.
Gene
2015
;
569
:
21
6
.
30.
Schmelzer
K
,
Fahy
E
,
Subramaniam
S
,
Dennis
EA
. 
The lipid maps initiative in lipidomics
.
Methods Enzymol
2007
;
432
:
171
83
.
31.
Tan
SLW
,
Chadha
S
,
Liu
Y
,
Gabasova
E
,
Perera
D
,
Ahmed
K
, et al
A class of environmental and endogenous toxins induces BRCA2 haploinsufficiency and genome instability
.
Cell
2017
;
169
:
1105
18
.
32.
Hatzivassiliou
G
,
Zhao
F
,
Bauer
DE
,
Andreadis
C
,
Shaw
AN
,
Dhanak
D
, et al
ATP citrate lyase inhibition can suppress tumor cell growth
.
Cancer Cell
2005
;
8
:
311
21
.
33.
Pirro
V
,
Alfaro
CM
,
Jarmusch
AK
,
Hattab
EM
,
Cohen-Gadol
AA
,
Cooks
RG
. 
Intraoperative assessment of tumor margins during glioma resection by desorption electrospray ionization-mass spectrometry
.
Proc Natl Acad Sci U S A
2017
;
114
:
6700
5
.
34.
Kandasamy
P
,
Numata
M
,
Berry
KZ
,
Fickes
R
,
Leslie
CC
,
Murphy
RC
, et al
Structural analogs of pulmonary surfactant phosphatidylglycerol inhibit toll-like receptor 2 and 4 signaling
.
J Lipid Res
2016
;
57
:
993
1005
.
35.
Furse
S
. 
Is phosphatidylglycerol essential for terrestrial life?
J Chem Biol
2016
;
10
:
1
9
.
36.
Wallace
DC
. 
Mitochondria and cancer
.
Nat Rev Cancer
2012
;
12
:
685
98
.
37.
Kiebish
MA
,
Han
X
,
Cheng
H
,
Chuang
JH
,
Seyfried
TN
. 
Cardiolipin and electron transport chain abnormalities in mouse brain tumor mitochondria: lipidomic evidence supporting the Warburg theory of cancer
.
J Lipid Res
2008
;
49
:
2545
56
.
38.
Schild
L
,
Lendeckel
U
,
Gardemann
A
,
Wiswedel
I
,
Schmidt
CA
,
Wolke
C
, et al
Composition of molecular cardiolipin species correlates with proliferation of lymphocytes
.
Exp Biol Med
2012
;
237
:
372
9
.
39.
Kim
J
,
Bowlby
R
,
Mungall
AJ
,
Robertson
AG
,
Odze
RD
,
Cherniack
AD
, et al
Integrated genomic characterization of oesophageal carcinoma
.
Nature
2017
;
541
:
169
75
.
40.
Segawa
K
,
Nagata
S
. 
An apoptotic ‘eat me' signal: phosphatidylserine exposure
.
Trends Cell Biol
2015
;
25
:
639
50
.
41.
Li
J
,
Condello
S
,
Thomes-Pepin
J
,
Ma
X
,
Xia
Y
,
Hurley
TD
, et al
Lipid desaturation is a metabolic marker and therapeutic target of ovarian cancer stem cells
.
Cell Stem Cell
2017
;
20
:
303
14
.
42.
Marien
E
,
Meister
M
,
Muley
T
,
Gomez Del Pulgar
T
,
Derua
R
,
Spraggins
JM
, et al
Phospholipid profiling identifies acyl chain elongation as a ubiquitous trait and potential target for the treatment of lung squamous cell carcinoma
.
Oncotarget
2016
;
7
:
12582
97
.
43.
Marien
E
,
Meister
M
,
Muley
T
,
Fieuws
S
,
Bordel
S
,
Derua
R
, et al
Non-small cell lung cancer is characterized by dramatic changes in phospholipid profiles
.
Int J Cancer
2015
;
137
:
1539
.
44.
Tamura
K
,
Makino
A
,
Hullin-Matsuda
F
,
Kobayashi
T
,
Furihata
M
,
Chung
S
, et al
Novel lipogenic enzyme ELOVL7 is involved in prostate cancer growth through saturated long-chain fatty acid metabolism
.
Cancer Res
2009
;
69
:
8133
40
.
45.
Yamashita
Y
,
Nishiumi
S
,
Kono
S
,
Takao
S
,
Azuma
T
,
Yoshida
M
. 
Differences in elongation of very long chain fatty acids and fatty acid metabolism between triple-negative and hormone receptor-positive breast cancer
.
BMC Cancer
2017
;
17
:
589
.
46.
Guri
Y
,
Colombi
M
,
Dazert
E
,
Riezman
I
,
Riezman
H
,
Hall
MN
. 
mTORC2 promotes tumorigenesis via lipid synthesis
.
Cancer Cell
2017
;
32
:
807
23
.
47.
Peck
B
,
Schulze
A
. 
Lipid desaturation - the next step in targeting lipogenesis in cancer?
FEBS J
2016
;
283
:
2767
78
.
48.
Flavin
R
,
Peluso
S
,
Nguyen
P
,
Loda
M
. 
Fatty acid synthase as a potential therapeutic target in cancer
.
Future Oncol
2010
;
6
:
551
62
.