Deep Metabolomics of a High-Grade Serous Ovarian Cancer Triple-Knockout Mouse Model

Abstract

High-grade serous carcinoma (HGSC) is the most common and deadliest ovarian cancer (OC) type, accounting for 70–80% of OC deaths. This high mortality is largely due to late diagnosis. Early detection is thus crucial to reduce mortality, yet the tumor pathogenesis of HGSC remains poorly understood, making early detection exceedingly difficult. Faithfully and reliably representing the clinical nature of human HGSC, a recently developed triple-knockout (TKO) mouse model offers a unique opportunity to examine the entire disease spectrum of HGSC. Metabolic alterations were investigated by applying ultra-performance liquid chromatography–mass spectrometry (UPLC–MS) to serum samples collected from these mice at premalignant, early, and advanced stages of HGSC. This comprehensive analysis revealed a panel of 29 serum metabolites that distinguished mice with HGSC from controls and mice with uterine tumors with over 95% accuracy. Meanwhile, our panel could further distinguish early-stage HGSC from controls with 100% accuracy and from advanced-stage HGSC with over 90% accuracy. Important identified metabolites included phospholipids, sphingomyelins, sterols, N-acyltaurine, oligopeptides, bilirubin, 2(3)-hydroxysebacic acids, uridine, N-acetylneuraminic acid, and pyrazine derivatives. Overall, our study provides insights into dysregulated metabolism associated with HGSC development and progression, and serves as a useful guide toward early detection.

Graphical Abstract

INTRODUCTION

Ovarian cancer (OC) is the most lethal gynecologic malignancy and the seventh most commonly diagnosed cancer among women in the world.¹ The overall 5-year survival rate of OC is 47%.² Due to the lack of specific symptoms, however, ovarian cancer is frequently diagnosed in late stages,^3,4 where 5-year survival rates are only 29%.² Among all the OC subtypes, high-grade serous carcinoma (HGSC), also known as high-grade serous ovarian cancer, is the most common and deadliest type, accounting for 70–80% of OC deaths.^2,5,6 The high mortality of HGSC is largely attributable to advanced-stage diagnosis. Approximately 80% of HGSC cases are diagnosed at advanced stage (stage III or IV), in which the 5-year and 10-year survival rates are 32.1% and 15%, respectively. Though only about 20% of HGSCs are detected at early stage (stage I or II), the early-stage diagnosis of HGSC dramatically raises patient survival to 71.4% (5 year) and 53% (10 year).^6,7 Hence, early detection of HGSC would be crucial to improving OC patients’ survival. Unfortunately, the cellular origin and tumor pathogenesis of HGSC still remain poorly understood, making early detection difficult.^5,8,9

Presently, no diagnostic tests are available for detecting HGSC at an early stage among at-risk patients, let alone screening tests for asymptomatic women in the general population.^6,10–13 The conventional strategy for OC risk evaluation includes trans-vaginal ultrasound and measurement of the serum tumor biomarker cancer antigen 125 (CA125) levels. However, this biomarker is of limited use, as serum elevations of CA125 can be observed in a number of different conditions unrelated to OC.¹⁴ In addition to CA125, there have been several biomarker-based tests approved by the FDA in ovarian cancer.¹⁵ However, these FDA-approved tests are neither diagnostic tests nor screening tests for OC. They are merely for referral purposes after ovarian tumor diagnosis is established. For example, the multivariate index assay OVA1 test is applied mainly to evaluate the likelihood of malignancy in women presenting an ovarian adnexal mass prior to surgery.^15,16

Metabolomics has emerged as a promising tool for biomarker discovery leading to enhanced diagnostics as well as providing insight into the molecular underpinnings of disease biology.¹⁷ Nuclear magnetic resonance (NMR) spectroscopy and mass spectrometry (MS) are the two techniques that have been most used for profiling metabolic alterations associated with OC in serum,^18–23 plasma,^24–27 urine,^28–31 tissue,^32–34 and ovarian cyst fluid.³⁵ Dysregulation in nucleotide, histidine, tryptophan, mucin,²⁸ phospholipid, and piperidine metabolic pathways,²⁴ fatty acid β-oxidation, and glycolysis³³ have been associated with OC development and progression. The majority of these new metabolic markers, however, are non-HGSC related.

One major impediment to HGSC early detection is the lack of knowledge of early tumor development and progression. To better understand the pathogenesis of ovarian cancer, we developed two mouse models of HGSC: (1) a double-knockout (DKO) mouse (Dicer1^flox/flox Pten^flox/flox Amhr2^cre/+) by inactivating the Dicer1 and Pten genes and (2) a triple mutant (TKO) mouse (p53^{LSL‑R172H/+} Dicer1^flox/flox Pten^flox/flox Amhr2^cre/+) by adding a p53 mutation (R172H), which is equivalent to human p53-R175H mutant, one of the frequent p53 mutations found in human OC.^36,37 As p53 mutations are observed in almost all human HGSC cases (96%),^38,39 the TKO mouse model would be genetically closer to human HGSC than the DKO model we previously studied.²⁰ These mice develop metastatic HGSC that faithfully model human HGSC phenotypically, histopathologically, and at the molecular level.³⁷ Remarkably, HGSC from these mice reproduces the clinical metastasis of human HGSC in 100% of cases. In these models, HGSC originates and develops progressively in the fallopian tube, envelops the ovaries, and then aggressively metastasizes throughout the peritoneal cavity, including the omentum, diaphragm, mesentery, and peritoneal membrane. These extensive metastases invariably induce massive ascites and inevitably kill the mice (100%). These mouse models therefore present a rare and unique opportunity to study the entire disease spectrum of HGSC—from inception to early progression to metastasis—particularly early-stage progression.

Here, using a “deep” ultra-performance liquid chromatography–mass spectrometry (UPLC–MS) approach, we characterize serum metabolic profiles of p53-Dicer1-Pten TKO mice at the premalignant stage, early stage, and advanced stage of HGSC. Among the identified are metabolites closely associated with HGSC, which discriminate HGSC from controls and from uterine tumors (UT) with over 95% accuracy, sensitivity, and specificity. The same metabolite panel is able to further distinguish early-stage HGSC from advanced-stage HGSC with over 90% accuracy, sensitivity, and specificity and distinguish early-stage HGSC against control samples with 100% sensitivity, specificity, and accuracy, respectively. The biological relevance of the differential metabolites is discussed, gaining insights into disease development and progression. This 29-feature panel enables effective detection of early stage as well as advanced-stage HGSCs, offering potential to diagnose HGSC at early stages.

MATERIALS AND METHODS

Chemicals

LC–MS-grade methanol, LC–MS-grade 2-propanol, LC–MS-grade acetonitrile, LC–MS-grade water, formic acid (99.5+%), ammonium acetate, and ammonium hydroxide were purchased from Fisher Chemical (Fisher Scientific International, Inc. Pittsburgh, PA) and used to prepare mobile phases and solutions.

Mice and Serum Sampling

In this study, triple mutant (TKO) mice (p53^{LSL‑R172H/+} Dicer1^flox/flox Pten^flox/flox Amhr2^cre/+) were generated by mating p53^{LSL‑R172H/+} Dicer1^flox/flox Pten^flox/flox female mice with Dicer1^flox/flox Pten^flox/flox Amhr2^cre/+ male mice. Serum samples were collected from TKO mice at different stages of tumor progression: from premalignant stage to early stage to advanced stage. Precursor lesions are in the fallopian tubes at the premalignant stage, when no tumors are yet present. TKO mice with precursor (i.e., premalignant stage) lesions in the fallopian tubes are referred to as “TKO-Pre”, TKO-ET (early tumor) are TKO mice with early stage tumors in the fallopian tubes without evidence of metastasis, TKO-AT (advanced stage tumor) are TKO mice with ovarian and peritoneal metastases accompanied by ascites, and TKO-Ctrl are control mice (p53^{LSL‑R172H/+} Dicer1^flox/flox Pten^flox/flox) that have the same genetic background as TKO mice but develop no tumors. All TKO mice developed high-grade serous carcinoma (HGSC, i.e., high-grade serous ovarian cancer). To enhance the selectivity for HGSC-specific metabolite markers, a tumor control group was also included. These were uterine tumor (UT) mice (p53^{LSL‑R172H/+} Pten^flox/flox Amhr2^cre/+) that developed uterine tumors, but not HGSC.

Blood samples were collected via retro-orbital bleeding after anesthesia, from 22 TKO-Pre mice (average age, 2.0 months; age range, 1.3–3.1 months), 10 TKO-ET mice (4.8 months, 2.5–5.9 months), 16 TKO-AT mice (6.4 months, 4.3–10.1 months), 19 TKO-Ctrl mice (3.0 months, 3.0–8.2 months), and 17 UT mice (5.4 months, 3.7–6.8 months). Blood samples were centrifuged at 14 000 rpm for 5 min at room temperature, and serum was collected and stored at −80 °C until UPLC–MS analysis.

Sample Preparation

Serum samples were thawed on ice and subject to two different sample preparation protocols to obtain profiles of both nonpolar and polar submetabolomes. Reverse phase (RP) and hydrophilic interaction liquid chromatography (HILIC) UPLC–MS analysis in both positive and negative ion modes were combined for these different polarity metabolite fractions to obtain complementary and “deeper” metabolome information. Methanol (for polar) or 2-propanol (for non-polar) was added to a 50 μL serum sample in a 3:1 ratio to precipitate proteins. Samples were vortex-mixed for 10 s and centrifuged at 13 000 rpm for 7 min. Then, 150 μL of supernatant was frozen at −80 °C for UPLC–MS analysis. A sample blank was prepared with 50 μL of LC–MS-grade water, and a pooled quality control (QC) sample was created by mixing a 10 μL aliquot of each serum sample. Both the sample blank and the pooled sample were processed with the same procedure as the murine serum samples. Samples were run in randomized order on consecutive days. Solvent blanks and sample blanks were analyzed together with murine serum samples. QC samples were analyzed every 12 runs to assess UPLC–MS system stability and were used to compensate for time-dependent batch effects with a QC-based regression curve for each detected compound.

UPLC–MS Analysis

Chromatography was performed with an Ultimate 3000 UPLC (Thermo Fisher Scientific, Inc., Waltham, MA) system equipped with a Waters ACQUITY UPLC BEH C18, 2.1 × 50 mm, 1.7 μm particle column or a Waters ACQUITY UPLC BEH HILIC, 2.1 × 75 mm, 1.7 μm particle column. A Q Exactive HF Orbitrap mass spectrometer (Thermo Fisher Scientific, Inc., Waltham, MA) was used in all cases. For reverse-phase (RP) separations, mobile phase A was water/ acetonitrile (40:60 v/v), and mobile phase B was acetonitrile/ 2-propanol (10:90 v/v). Both mobile phases included 10 mM ammonium formate and 0.1% formic acid additives to improve peak shape and ionization efficiency. For hydrophilic interaction chromatography (HILIC) separations, mobile phase A was water/acetonitrile (95:5 v/v), 10 mM ammonium acetate, and 0.05% ammonium hydroxide. Mobile phase B was acetonitrile with 0.05% ammonium hydroxide. Chromatographic gradients are described in Table S1 of the Supporting Information (SI). The column temperature was 55 °C, while samples were maintained at 5 °C in the autosampler. Injection volumes of 5 and 2 μL were used in RP and HILIC methods, respectively. RP and HILIC chromatographies were performed both in positive and negative ion modes, and MS parameters can be found in the Supporting Information (Table S2).

For metabolite identification purposes, the top five data-dependent acquisition (DDA) experiments were used to collect MS/MS spectra using a stepped normalized collision energy (NCE) of 10, 30, and 50 V. For compounds missed by DDA, parallel reaction monitoring (PRM) experiments were performed at collision energies ranging from 10 to 40 V to obtain fragmentation data for identification purposes.

Data Processing

Spectral features were extracted from the raw data using Compound Discoverer v2.1 software (Thermo Fisher Scientific, Inc., Waltham, MA). This procedure included chromatographic alignment, peak picking, peak area integration, and QC-based compound area normalization. Features that eluted with the chromatographic solvent front with retention times <0.5 min in RP data sets and <0.9 min in HILIC data sets were considered unreliable due to potential ion suppression effects⁴⁰ and were thus removed. Further filtering was carried out by removing features that were not present in 50% of at least one of the serum sample groups at 10 times the baseline abundance, defined as the peak area of the sample blank run. Welch’s t test with a Benjamini Hochberg correction was applied to TKO vs TKO-Ctrl mouse groups and to TKO vs UT mouse groups. The union of differential features from both control and UT comparisons was chosen for down-selecting HGSC-specific features. The most useful set of features for classification was produced with genetic algorithms⁴¹ (GAs, MATLAB R2016a, The MathWorks, Natick, MA, with PLS_Toolbox v.8.1.1, Eigenvector Research, Inc., Wenatchee, WA). GAs are evolutionary algorithms that generate solutions to optimization problems. They have been widely used in variable selection due to their high performance in large-scale variable selection rate and classification accuracy.⁴¹ Parameters for GA variable selection are provided in the Supporting Information (Table S3). After the GA variable selection process, features were chosen on the basis of frequency criteria. Principal component analysis (PCA) and orthogonal partial least square discriminant analysis (oPLS-DA)⁴² were performed to assess the discriminating power of metabolite data sets in nonsupervised and supervised manners, respectively. Data were preprocessed by autoscaling prior to PCA and oPLS-DA analysis and cross-validated using 10 iterations of random sample subsets. Data generated in this work are available through the NIH Metabolomics Workbench (http://www.metabolomicsworkbench.org/) with project ID PR000784 (doi 10.21228/M8BH6F, study ID ST001172).

Discriminant Feature Identification

Metabolite identification was attempted for the panel of best discriminant features. Mass spectral ion adduct analysis was first performed to ensure the unambiguous assignment of the feature of interest in each mass spectrum. Elemental formulas were then generated on the basis of exact masses with a maximum mass error of 10 mDa and isotopic patterns using Compound Discoverer v2.1. The elements included in the elemental formula search were constrained to C, H, N, O, P, and S. Tentative identities were searched against the human metabolome database (HMDB),⁴³ Lipid Maps database,⁴⁴ and Metlin database⁴⁵ using both the accurate mass and generated elemental formulas with a mass error of 10 mDa. Tandem MS databases such as Metlin, mzCloud,⁴⁶ and MassBank⁴⁷ were used together with literature searches to further confirm the identity of the metabolite candidates for which tandem MS/ MS data were successfully acquired. Fragmentation patterns were also manually analyzed in a few cases to distinguish between different isobaric species.

RESULTS AND DISCUSSION

Multivariate Classification Performance

A data set comprising a total of 5937 spectral features that were above background and chromatographically retained was produced by combining the RP ESI(+), RP ESI(−), HILIC ESI(+), and HILIC ESI(−) data sets. Initially, the extent by which UPLC–MS metabolic profiling could differentiate serum samples of 19 TKO-Pre, 10 TKO-ET, and 16 TKO-AT mice, against 22 TKO-ctrl and 17 UT mice was investigated. An unsupervised PCA exploration of these data sets showed large overlap between TKO-Pre and TKO-Ctrl mice (Figure 1), indicating that, as expected, only subtle differences existed between the two groups because TKO-Pre mice had not yet developed any tumors and were phenotypically closer to the healthy controls. Thus, for the following data analysis, the TKO mouse group is a combination of TKO-ET and TKO-AT mice.

Figure 1.

PCA score plot including all sample groups using the total initial set of 5937 spectral features. The clustering of samples in this plot reveals clear separation between TKO-ET mice from TKO-AT mice and UT mice, a moderate separation between TKO-ET mice and TKO-Ctrl mice with some overlap, and complete overlap of TKO-Pre and TKO-Ctrl mice. Pooled QC samples, represented by orange stars, clustered toward the center of the plot, indicating that the technical variance was minimal.

The set of 5937 features was utilized to build an oPLS-DA model that distinguished serum samples from TKO mice against TKO-Ctrl and UT mice. Performance characteristics of this model (Figure 2A,B) were 88.1, 92.8, and 90.4% for the cross-validated sensitivity, specificity, and accuracy, respectively. Five serum samples were systematically misclassified, including two TKO-ET mice with ID No. 42 and 51, one TKO-AT mouse labeled No. 62, and two UT mice labeled No. 69 and 78. Sample cohort information can be found in the Supporting Information. The model employed a total of four latent variables and interpreted 42.1% and 54.48% variance from the X (feature peak area) and Y (sample class membership) blocks, respectively. Although this model’s performance was acceptable, GA feature selection was used to obtain a smaller, more robust, metabolic feature set that could better discriminate TKO mice from TKO-Ctrl and UT mice. GA variable selection led to a 29-feature panel (Figure 2C,D) with 96.2, 97.2, and 96.7% cross-validated sensitivity, specificity, and accuracy, respectively. In this case, only two serum samples were systematically misclassified. The two misclassified samples were collected from a TKO-ET mouse labeled No. 42 and a UT mouse labeled No. 68. The TKO-ET mouse was repeatedly misclassified as a TKO-Ctrl mouse. This could be due to the age of the mouse, 2.5 months, when the sample was collected, which was much younger than the rest of the TKO-ET mice (average age 4.8 months) and close to the average age of the TKO-Pre mice (average age 2.0 months). As TKO-Pre mice were phenotypically closer to the TKO-Ctrl mice, this may explain why the young TKO-ET mouse was classified as a TKO-Ctrl mouse in the models. The three latent variable model interpreted 58.7% and 56.1% variance from the X and Y blocks, respectively. Binary comparisons (TKO vs TKO-Ctrl, TKO vs UT) were also performed (Figure 3A,B), indicating that the 29-feature panel was strongly HGSC-specific and could differentiate TKO mice from control and UT mice with high performance at the same time. In these cases, the classification accuracy ranged from 95.3 to 98.1%, with a cross-validated sensitivity of 96.2–96.5% and specificity of 94.1–100%. PCA was utilized to further evaluate the discriminant performance of the 29-feature panel in an unsupervised manner and rule out any overfitting. Score plots were generated for both the initial 5937-feature set and the selected 29-feature panel (Figure S1, SI). Good group clustering and increased variance captured were observed with the 29-feature panel, even in an unsupervised fashion.

Figure 2.

(A) oPLS-DA score plot depicting clustering of samples using the initial set of 5937 spectral features. (B) oPLS-DA cross-validated classification plot using 5937 spectral features. The x-axis represents randomized sample number, and the y-axis represents the cross-validated predicted scores of the oPLS-DA classification model. (C) oPLS-DA scores plot using the GA-selected 29-feature panel. (D) oPLS-DA cross-validated classification plot using the 29-feature panel. TKO, TKO-Ctrl, and UT samples are represented by black squares, red circles, and blue triangles, respectively. The threshold for sample classification is represented by the green dashed line. Cross-validated sensitivity, specificity, and accuracy values are given for each model. TKO samples are a combination of TKO-ET and TKO-AT samples.

Figure 3.

oPLS-DA scores plot depicting clustering of samples between (A) TKO and TKO-Ctrl samples, (B) TKO and UT samples, and (C) TKO-ET and TKO-AT samples using the 29-feature panel. Variance between classes is captured across the x-axis. TKO, TKO-Ctrl, UT, TKO-ET, and TKO-AT samples are represented by black squares, red circles, blue triangles, green diamonds, and pink triangles, respectively. Cross-validated sensitivity, specificity, and accuracy values are given for each model. TKO samples are TKO-ET and TKO-AT samples combined.

As accurate staging and early-stage-disease detection greatly improve clinical outcome, we investigated the performance of the 29-feature panel in discriminating early and late stages of HGSC. This panel of 29 metabolites distinguished TKO-ET from TKO-AT mice with 90.0%, 93.8%, and 91.9% for the cross-validated sensitivity, specificity, and accuracy, respectively (Figure 3C). Permutation tests with 2000 iterations returned a p-value of 0.017 measured using group separation distance,⁴⁸ avoided overfitting, and further validated the model. We further explored the potential of the 29-feature panel in disease early detection. The oPLS-DA model successfully discriminated TKO-ET against TKO-Ctrl mice with 100% cross-validated sensitivity, specificity, and accuracy, respectively. This panel, however, was indistinguishable between precursors and controls. We used the 22 precursor (TKO-Pre) samples as an unknown sample set, and input into this classification model, 18 out of 22 TKO-Pre samples were predicted as being similar to TKO-Ctrl samples (Figure 4), echoing our previous statement that the TKO-Pre mice were phenotypically closer to the TKO-Ctrl mice. Nevertheless, this 29-feature panel enabled effective detection of early stage as well as advanced-stage HGSCs, offering potential to diagnose HGSC at early stage.

Figure 4.

oPLS-DA cross-validated classification plot between ET and Ctrl samples using the 29-feature panel. The x-axis represents the randomized sample number, and the y-axis represents the cross-validated predicted scores of the binary oPLS-DA classification model. ET and Ctrl samples are represented by green diamonds and red circles, respectively. Precursor (Pre) samples projected into the model are represented by orange triangles. The threshold for sample classification is represented by the green dashed line. Cross-validated sensitivity, specificity, and accuracy values are given for the model.

Discriminant Metabolite Identification

Following multivariate analysis, metabolite identification was carried out for the species in the 29-feature panel. Twenty-four of the 29 metabolic features were identified by both high-resolution MS and MS/MS (Table 1). Table S4 (SI) provides detailed MS/MS fragmentation information and confidence level for each identified species. Metabolites in the panel included lipids, oligopeptides, and other small molecules, such as bilirubin, uridine, hydroxysebacic acids, and N-acetylneuraminic acid (NeuAc). For features that did not yield informative MS/MS information, tentative assignments were made on the basis of high-resolution MS and isotopic relative ion abundances only (Table 1, italicized) and should be considered tentative until further research is pursued.

Table 1.

Annotation of Metabolites in the 29-Feature Panel, Including the Acquisition Mode, Retention Time, Observed Exact Mass and Mass Error, Theoretical Mass, Predicted Elemental Formula, Observed Adduct, p-Value of Abundances between TKO and TKO-Ctrl Samples and between TKO and UT Samples, and Fold Changes

			m/z
Feature no.	Acquisition mode	retention time (min)	expl	theor	Elemental formula	adduct type	mass error (ppm)	Fold changea	p-valueb HGSC vs control/HGSC vs UT	metabolite identityc
208	HILIC ESI(+)	4.96	109.0765	109.0760	C6H8N2	[M + H]+	4.58	−0.19	0.003 47/8.45 × 10−7	dimethypyrazine, ethylpyrazine, 2-picolylamine
202	HILIC ESI(+)	2.67	271.1151	271.1149	C9H14N6O4	[M + H]+	0.74	1.08	0.000 261/1.52 × 10−6	–
385	HILIC ESI(−)	1.65	516.7524	–	–	[M + Br]−	–	0.68	0.001 16/2.86 × 10−6	–
277	RP ESI(−)	1.02	583.2576	583.2562	C33H36N4O6	[M − H]−	2.40	−0.88	1.10 × 10−7/5.57 × 10−6	bilirubin
494	HILIC ESI(−)	1.58	514.7545	–	–	[M + Br]−	–	1.33	0.000 762/2.51 × 10−5	–
363	HILIC ESI(+)	2.21	398.2022	398.2034	C17H27N5O6	[M + H]+	−3.01	0.80	0.001 48/3.13 × 10−5	Glu-His-Leu Glu-His-He
427	RP ESI(+)	8.15	693.5567	693.5581	C47H74O2	[M + Na]+	−2.02	1.26	0.001 04/0.000 112	CE(20:5)
303	RP ESI(+)	1.42	552.4015	552.4024	C28H58NO7P	[M + H]+	−1.63	0.93	0.005 46/0.000 114	LysoPC(20:0)
751	HILIC ESI(+)	1.28	696.4993	696.5012	C76H144O17P2	[M + 2H]2+	−2.73	0.57	0.0116/0.000 124	CL(67:2)
37	HILIC ESI(−)	2.20	217.1075	217.1081	C10H18O5	[M − H]−	−2.76	−0.42	0.006 97/0.000 137	2(3)-hydroxysebacic acid
572	HILIC ESI(+)	4.56	454.8905	–	–	[M + H]+	–	−1.50	9.44 × 10−7/0.000 149	–
651	HILIC ESI(−)	1.27	281.0358	281.0360	C9H12N2O6	[M + Cl]−	−0.71	−0.22	0.009 34/0.000 351	uridine (isotopic peak)
557	HILIC ESI(−)	3.22	308.0988	308.0987	C11H19NO9	[M − H]−	0.32	0.94	0.003 16/0.000 360	NeuAc
88	RP ESI(+)	1.22	607.2522	607.2527	C33H36N4O6	[M + Na]+	−0.82	−1.10	1.34 × 10−6/0.000 491	bilirubin
274	RP ESI(+)	1.25	540.4465	540.4445	C32H63NO4S	[M + H − H20]+	3.70	−0.84	3.19 × 10−6/0.000 718	NAT(30:1)
696	HILIC ESI(−)	1.27	279.0387	279.0389	C9H12N2O6	[M + Cl]−	−0.72	−2.21	9.20 × 10−14/0.000 865	uridine
449	RP ESI(−)	6.88	673.5272	673.5290	C37H75N2O6P	[M − H]−	−2.67	−1.75	0.000 209/0.001 21	SM(d32:1)
189	RP ESI(+)	1.20	585.2700	585.2708	C33H36N4O6	[M + H]+	−1.37	−0.21	0.0161/0.001 52	bilirubin
210	HILIC ESI(+)	4.31	619.1976	–	–	[M + H]+	–	0.18	0.009 58/0.002 14	–
98	HILIC ESI(−)	2.68	247.1184	247.1187	C11H20O6	[M − H]−	−1.21	−1.98	3.57 × 10−10/0.002 37	fatty acyl glucoside
191	RP ESI(+)	4.47	818.6049	818.6058	C48H84NO7P	[M + H]+	−1.10	−0.36	0.0145/0.003 84	PC(P-40:6)
485	RP ESI(+)	6.27	978.7103	978.7158	C56H100NO10P	[M + H]+	−5.62	0.16	0.000 306/0.004 13	PS(50:5), PS(O-50:6(OH)), PS(P-50:5(OH))
700	RP ESI(−)	1.06	472.1559	472.1586	C20H23N7O7	[M − H]−	−5.72	0.75	0.001 16/0.006 80	folinic acid, 10-formyltetrahydrofolate, pteroyl-d-glutamic acid
22	RP ESI(−)	3.70	491.3419	491.3412	C26H52O6S	[M − H]−	1.42	3.03	0.002 25/0.007 11	fatty acid ester derivatives
644	HILIC ESI(+)	3.93	276.1190	276.1190	C10H17N3O6	[M + H]+	0.00	2.93	0.002 93/0.008 67	norophthalmic acid, γ-glutamyl glutamine
725	HILIC ESI(−)	1.25	219.0187	219.0193	C5H8N4O4S	[M − H]−	−2.74	−0.29	0.0120/0.008 90	taurine derivatives
48	RP ESI(−)	0.93	462.0594	462.0616	C14H25NO11	[M + Br]−	−4.76	2.78	0.002 69/0.009 63	N-acetyllactosamine, β-1,4-mannose-N-acetylglucosamine, lacto-N-biose I
772	HILIC ESI(−)	3.61	217.1075	217.1081	C10H18O5	[M − H]−	−2.76	−0.90	0.001 14/0.0117	2(3)-hydroxysebacic acid
29	RP ESI(+)	3.27	383.3306	383.3308	C27H42O	[M + H]+	−0.52	−1.06	1.25 × 10−9/0.0127	7-dehydrodesmosterol, 5a-cholesta-8,24-dien-3-one, cholesta-4,6-dien-3-one

^aFold change (FC) was calculated as the base 2 logarithm of the average abundance ratios between TKO and TKO-Ctrl samples). Positive FC values indicate increased abundance in TKO samples, while negative values indicate higher abundance in TKO-Ctrl samples.

^bAll p-values are calculated using an FDR-corrected t-test. For additional fragmentation information and level of confidence for each identification, please refer to Table S4 (SI).

^cAbbreviations: Glu, glutamic acid; His, histidine; Leu, leucine; Ile, isoleucine; CE, cholesteryl ester; LysoPC, lysophosphatidylcholine; PC, phosphatidylcholine; CL, cardiolipin; NAT, N-acyltaurine; SM, sphingomyelin; PS, phosphatidylserine; NeuAc, N-acetylneuraminic acid.

HGSC-Related Metabolic Alterations

Understanding the biological role of specific metabolites is also crucial to enhance our understanding of HGSC metabolism. Discussed below is the potential role of the differential metabolites that were altered in different HGSC stages (Ctrl, Pre, ET, and AT). The biological roles of the remaining identified metabolites can be found in the Supporting Information.

Alterations in metabolism of phosphatidylcholine (PC) and its ester-bond hydrolysis product, lysophosphatidylcholine (LysoPC), have been reported in several gynecological cancers, including ovarian cancer studies,^49–53 cervical cancer,⁵⁴ and breast cancer.⁵⁵ Our findings of decreased PC(P-40:6) (Table 1; fold change −0.36) and elevated LysoPC(20:0) (Table 1; fold change +0.93) serum levels in TKO-AT mice are in agreement with published ovarian cancer studies.^51,53 The alterations are likely due to the activation of PC-cycle enzymes, including choline kinase (ChoK) and PC-specific phospholipase C (PC-plc).^50,56 Moreover, the significantly decreased PC and elevated LysoPC alterations were found in advanced-stage cancer (Figure 5), suggesting that these changes were related to tumor metastasis and progression. This observation awaits further validation and investigation.

Figure 5.

Box plots showing changes of the selected identified metabolites in Ctrl (n = 19), Pre (n = 22), ET (n = 10), and AT (n = 16) samples. The mean, median, upper and lower quartiles, outliers, and minimum and maximum (whiskers) values are displayed. *Features have isomers.

Cardiolipins (CLs) are a unique mitochondrial phospholipid class that regulates bioenergetic processes and signaling events related to apoptosis and aging.^57,58 An increase of CL(67:2) serum levels in TKO-ET mice was observed (Figure 5). However, further studies are still needed to understand the role of CL in cellular function and signaling pathways as they relate to ovarian cancer.

One of the identified features in the 29-metabolite panel was the sphingomyelin (SM) SM(d32:1) (Table 1). SM is an essential element of plasma membrane structure and plays important roles in cancer biology, with the hydrolysis by sphingomyelinases resulting in the formation of ceramides, key players in cellular proliferation, growth, and apoptosis.^59,60 Consequently, changes in SM levels can have a profound effect on the biophysical properties of cellular membranes and signaling.⁶¹ We observed decreased levels of SM(d32:1) in HGSC mice, especially in the advanced stage (Table 1, Figure 5; fold change −1.75). Along the same lines, Braicu et al. have detected a decrease of SM levels in serum samples from ovarian cancer patients,⁶² strengthening the findings of our study.

Sterol metabolites derived from the cholesterol biosynthetic pathway are important structural components of cell membranes.⁶³ They regulate membrane fluidity and permeability as well as biological activities, including lipid synthesis, cell growth, and apoptosis.⁶⁴ One discriminant feature was identified as a sterol metabolite, cholesteryl ester [CE(20:5)]. CEs are formed by the esterification of cholesterol with long-chain fatty acids and are transported through the blood by lipoproteins. Overexpression of CEs has been reported in multiple malignant tumors,⁶⁵ including ovarian cancer.⁶⁶ We observed a significant increase in serum levels of CE(20:5) in TKO-AT mice (Table 1, Figure 5; fold change +1.26), which is in line with these findings. Another cholesterol derivative, C₂₇H₄₃O [7-dehydrodesmosterol, 5a-cholesta-8,24-dien-3-one, or cholesta-4,6-dien-3-one], was identified in our 29-feature panel. Gradually decreased serum levels of 7-dehydrodesmosterol (and/or its isomers) were detected in TKO with different stages (Table 1, Figure 5; fold change −1.06). 7-Dehydrodesmosterol (and/or its isomers) is an intermediate in the cholesterol biosynthesis pathway. The decrease can be attributed to the enhanced biosynthesis activity of cholesterol, as an increased serum cholesterol level is associated with ovarian cancer.⁶⁷

N-Acyltaurines (NATs) are molecules with fatty acids conjugated to taurine. We observed decreased serum levels of NAT(30:1) in HGSC mice, especially in the advanced stage (Table 1, Figure 5; fold change −0.84). Limited research has been reported on these species, and their biological role remains largely unexplored. In support of our findings, Chatzakos et al. found that NAT reduced proliferation in human prostate cancer cells.⁶⁸ Therefore, one likely explanation is the diminished antiproliferative effect with decreased levels of NAT, resulting in further disease progression.

A tripeptide (Glu-His-Leu or Glu-His-Ile) was found to be higher in TKO-ET mice (Table 1, Figure 5; fold change +0.80). Literature searches revealed no known roles for small peptides in ovarian cancer. Some proteases such as protease M, however, have been reported to be overexpressed in serum of ovarian cancer patients^69–72 and could, therefore, increase the levels of small circulating peptides in HGSC.

Bilirubin, the end product of heme catabolism in mammals, exhibits antioxidant properties by scavenging peroxyl radicals.^73,74 Increased serum bilirubin concentration has been found to be associated with decreased risk for cancer mortality.⁷⁵ We observed decreased serum levels of bilirubin in HGSC mice, with further decrease in the advanced stage (Table 1, Figure 5; fold change [M − H]⁻ −0.88). This finding agrees with previous work by our group, showing lowered concentration of bilirubin for early-stage ovarian cancer in Dicer-Pten double-knockout (DKO) mice compared with controls.²⁰ One likely explanation of this finding is the diminished protection of bilirubin against oxidative stress in ovarian cancer, leading to further disease progression.

2(3)-Hydroxysebacic acids have been found to be important diagnostic markers for impaired peroxisomal fatty acid oxidation disorders.⁷⁶ Their levels are regulated by medium-chain acyl-CoA dehydrogenase in the fatty acid β-oxidation pathway. Increased excretion of 2(3)-hydroxysebacic acids is associated with acyl-CoA dehydrogenase deficiency and decreased β-oxidation of fatty acids.^76–78 Our experiments reported lower serum levels of 2(3)-hydroxysebacic acids in TKO-AT mice (Table 1, Figure 5; fold change −0.90), suggesting increased activity of acyl-CoA dehydrogenase and fatty acid β-oxidation. In support of this hypothesis, both Fong et al. and Ke et al. have observed increased fatty acid β-oxidation in ovarian cancer patients.^24,33

We observed significantly decreased serum levels of uridine in HGSC mice, especially in the early stage (Table 1, Figure 5; fold change −2.21). The level of circulating uridine is a reflection of de novo pyrimidine biosynthesis and the utilization of uridine by tissue via the salvage pathway.⁷⁹ As cell proliferation in tumors requires a large pool of pyrimidines for rapid DNA and RNA synthesis, the observed relative decrease in serum uridine levels could be due to either tissue uridine uptake or rapid uridine catabolism in ovarian cancer.

N-Acetylneuraminic acid (NeuAc), also known as sialic acid (SA), is a major constituent of glycoproteins and glycolipids.⁸⁰ We observed significantly increased serum NeuAc levels in TKO-AT mice (Table 1, Figure 5; fold change 0.94). In agreement with this finding, increased SA serum levels have been reported in ovarian cancer patients compared to healthy controls.^81,82 This elevation of NeuAc serum levels could be attributed to high levels of glycoprotein and glycolipid release due to the high turnover of malignant tumor cells.^80,83 Moreover, the detected elevation of NeuAc levels in TKO-AT mice suggests its strong association with disease progression, confirming its potential application as a diagnostic indicator for HGSC malignancy.^81,84,85

One feature in the discriminant metabolite panel was identified as C₆H₈N₂, likely a pyrazine derivative, with lower abundance in TKO-AT mice (Table 1, Figure 5; fold change −0.19). Many pyrazines are exogenous metabolites that originate from food sources, such as cereals, soybeans, potatoes, cocoa products, and other natural products,⁸⁶ and have been reported to exhibit anticancer activities.⁸⁷ Although all mice in this study were fed the same diet, it is plausible that altered metabolism in HGSC mice led to differences in abundance of dietary metabolites following disease progression.

CONCLUSIONS

Serum from TKO p53-Dicer1-Pten mice with both early- and advanced-stage HGSC tumors was successfully profiled using a UPLC–MS-based, nontargeted metabolomics strategy. Identified are a panel of 29 metabolites that distinguishes TKO mice from TKO-Ctrl and UT mice and further distinguishes TKO-ET from TKO-AT and TKO-Ctrl. Metabolic alterations in TKO mice—mainly characterized by aggressive fatty acid β-oxidation; abnormal metabolism of phospholipids, glycoproteins, and glycolipids; heme catabolism; cholesterol biosynthesis; and pyrimidine biosynthesis—were found to be associated with HGSC development and progression. Identified biomarkers that differentially expressed in early and advanced stages of HGSC were phospholipids, sphingomyelin, sterols, N-acyltaurine, oligopeptide, and other small molecules, including bilirubin, 2(3)-hydroxysebacic acids, uridine, N-acetylneuraminic acid, and pyrazine derivatives. Taken together, our deep metabolomics study provided insights into dysregulated metabolism in HGSC, which could aid in disease diagnosis, as well as support our understanding of disease development and progression. In particular, this panel of 29 metabolites will serve as a useful guide toward early detection of high-grade serous ovarian cancer.