Integrative Systemic and Local Metabolomics with Impact on Survival in High-Grade Serous Ovarian Cancer

Metabolomic changes (“deregulating cellular energetics”) of cancer cells were added to “the next generation” hallmarks of cancer in 2011 (1), but anomalous characteristic of cancer cell energy metabolism, i.e., “aerobic glycolysis,” was already observed by Otto Warburg back in 1930. Little is known about how and wherefrom cancer cells get fuel, building blocks, and reducing equivalents for their increased cell growth and division. Limiting factors are oxygen, oxidizable fuel for oxidative phosphorylation in mitochondria and aerobic glycolysis, building blocks for proteins, i.e., essential amino acids (AA), and especially for hypoxic cancer tissues: unsaturated fatty acids (desaturation is an oxygen-dependent process). Especially, increased consumption of glucose (for energy production) and glutamine (as carbon and reduced nitrogen source) by tumors in comparison with the nonproliferating normal tissues is evident (2). Phospholipids can serve as storage of energy in the form of fatty acyl chains, utilized only under specific conditions such as starvation (3) and possibly also by energy-craving malignant cells.

Epithelial ovarian cancer (EOC) is characterized by a very unfavorable outcome, mainly due to late diagnosis accompanied by a high tumor burden at first presentation. Local tumor spread and local tumor recurrence of finally chemotherapy-resistant tumors in the peritoneal cavity and associated clinical complications are the main causes of lethality.

Aim of this study was to find survival relevant systemic metabolic changes in cancer patients and to define connections of these changes to the local microenvironment, i.e., the ascites and the tumor tissue itself, and to uncover systemic and local immune system and tumor cell characteristics associated with these metabolic changes. Metabolic changes and heterogeneity in tumor tissues compared with adjacent normal tissues were subject of comprehensive analyses recently (4, 5), but little is known if and how changes in serum metabolites may predict outcome in cancer patients, even less if an increased consumption of fuels and building blocks by tumors is correlated with a measurable reduction of metabolites in blood at all. Ovarian cancer seems to be a good model-malignancy to study systemic metabolomic changes because of its high tumor burden at first diagnosis, complete reduction of the tumor mass after first therapy in the majority of patients, and similar causes of lethality, i.e., complications associated with tumor masses in the peritoneal cavity. To avoid histologic and molecular biases, only late-stage, i.e., FIGO III/IV, high-grade serous ovarian cancer (HGSOC) patients were used for a matched serum, ascites, and tumor tissue metabolomics study. The impact of serum metabolite changes on overall survival (OS) was validated via a gene signature in an independent patient cohort.

Serum of 65 therapy-naïve HGSOC patients and 62 healthy controls was used. Ascites and fresh tumor samples were collected and processed as previously described (ref. 6; Supplementary Fig. S1; Supplementary Table S1).

Targeted metabolomics was performed with AbsoluteIDQ p180 kits (Biocrates Life Sciences AG) after the guidelines from Biocrates.

Vesicle-free soluble fractions and pellets were isolated from mixtures of each five sera from cancer patients or healthy controls, and human high-density lipoproteins (HDL), low-density lipoproteins (LDL), and very low-density lipoproteins (vLDL) were obtained from Sigma-Aldrich. Exosomes were isolated from sera mixtures.

Total cholesterol, triglycerides, LDL, HDL, albumin, and C-reactive protein (CRP) were measured on a cobas 8000 modular analyzer (Hoffmann-LaRoche). Luminex-based analyses were performed from cell free ascites and serum following instructions provided by the corresponding kits on a Bio-Plex 200 System (Bio-Rad Laboratories): “Bio-Plex Pro Human Cancer Biomarker Assays: Panel 1” (n = 16), “Bio-Plex Pro Human Chemokine Panel Assay” (n = 40; both Bio-Rad Laboratories), and “Cytokine Human Magnetic 25-Plex Panel” (Life Technologies).

Staining was performed on formalin-fixed and paraffin-embedded tumor tissue sections from ovarian tumor masses (P, “primary” tumors) and peritoneal tumor masses (M, “metastases”) with a Leica BOND-III Fully Automated IHC & ISH system (Leica Biosystems Nussloch GmbH) using PD-L1 (E1L3N) XP Rabbit mAb (Cell Signaling Technology), PD-1 Cell Marque 315M-96 monoclonal mouse anti-human antibody (Sigma-Aldrich), and CD8 (clone C8/144B), DAKO CD8 M7103 monoclonal mouse anti-human antibody (Agilent Technologies).

RNA sequencing was performed as described in ref. 6.

Survival analyses were performed by univariate and multiple Cox-regression analyses using clinicopathologic factors and relevant laboratory measures as covariates (see Table 1). The glycerophospholipid (GPhL) score was trichotomized along the 33.33 and 66.67 percentiles.

For validation of the impact of the GPhL score on OS, the 100-gene expression signature was used as surrogate marker for the GPhL score and validated in an independent cohort of serous ovarian cancer patients (6–9). The gene predictor was calculated as follows: median 50 top positively correlated genes minus median 50 top negatively correlated genes.

From 65 HGSOC patients or subsets thereof, pretreatment and follow-up sera, pretreatment ascites (the microenvironment), and pretreatment tumor tissues were analyzed for metabolic changes in association with OS and with each other. A comprehensive description of samples and methods is given in Supplementary Fig. S1 and Supplementary Table S1. The metabolic impact on OS was compared with and corrected for usual clinicopathologic and laboratory predictors. Sera of 62 individuals, either healthy or patients with nonmalignant gynecologic diseases, were used as controls. To uncover connections to the systemic and local immune system, classical inflammatory markers and cyto/chemokines were measured from blood and ascites, and the level and characteristic of tumor-infiltrating immune cells were assessed by analysis of the immune checkpoint system PD-1 and PD-L1. The interrelations between systemic and local metabolic changes and the tumor transcriptomes were assessed by RNA sequencing of isolated tumor cells and by comparison with established molecular subclasses for HGSOC. A gene signature predictive for OS-associated metabolic changes was used to validate the impact on OS in an independent patient cohort.

From sera of 65 untreated HGSOC patients and 62 healthy controls, 179 metabolites from six metabolite classes, GPhLs, sphingolipids, acylcarnitines, AAs, biogenic amines, and the sum of hexoses (mainly glucose), were measured by targeted metabolomics. Correlations between median concentrations of metabolite classes were substantial (r ≥ 0.40) between AAs and biogenic amines, between both classes of lipids, and interestingly also between AAs and GPhLs (Supplementary Fig. S2).

A robust (permutation bases thus non-parametric) method with the censored OS data as outcome was used to determine the serum metabolites which were associated significantly with survival outcome (Fig. 1A). Forty-three GPhLs of 89 (43.9%), 5 AAs of 21 (23.8%, four proteinogenic: histidine, lysine, threonine, tryptophan; and citrulline, a byproduct of the enzymatic production of nitric oxide and a marker—if reduced—for severe inflammation and sepsis in children; ref. 10), 4 sphingolipids of 15 (26.7%), and 1 biogenic amine of 13 (7.7%, acetylornithine, which is involved in glutamate biosynthesis) correlated significantly negatively with OS, i.e., low concentration with unfavorable survival (Fig. 1A). Significant OS-associated GPhLs and AAs were decreased in patients with short OS (<3 years) compared with healthy controls but similar, or in the case of AAs only slightly decreased, to controls in patients with long OS (>3 years). Patients with unknown 3-year survival (alive, but follow-up shorter than 3 years) were—as expected—between these two groups (Fig. 1B). The median concentrations of nonsignificant AAs were similar for all groups (Fig. 1B). In Supplementary Fig. S3, median concentrations of nonsignificant GPhLs, all analyzed GPhLs, other metabolite classes, and selected single metabolites are shown. Interestingly, glutamine, known to be consumed by tumors, was reduced neither in cancer patients compared with healthy controls nor in patients with shorter compared with longer OS. The sum of hexoses, i.e., in blood more than 90% glucose the remaining fructose, was even significantly elevated in cancer patients compared with controls, independent of OS (Supplementary Fig. S3). In Supplementary Fig. S4, log₂ concentrations of all GPhLs are shown, subdivided into healthy controls and correspondingly the 4 patients with the most extreme (highest and lowest) GPhL levels.

All four significantly OS-associated proteinogenic AAs are (nonconditional) essential for humans, which is a significant enrichment compared with non-essential AAs (P = 0.033; Supplementary Fig. S5). A detailed analysis of GPhLs revealed that mainly medium-length (sum of both fatty acids from 32 to 42 carbons) and (poly)unsaturated phospholipids (steadily increasing up to an unsaturation degree of three) of both types, O-acyl-O-acyl and O-alkyl-O-acyl, were significantly associated with OS (Fig. 1C). In Supplementary Fig. S6, median concentrations of significant OS-associated GPhLs are shown, split according to the degree of unsaturation in controls and patients with different 3-year survivals. Four- and 5-time unsaturated GPhLs by far showed the highest concentrations in serum, and 6-time unsaturated GPhLs showed the lowest concentrations. But the most important OS predictors (in number and impact) are the medium concentrated 1- to 3-time unsaturated GPhLs (cf. Fig. 1C). For example, 10 of 12 (83%) 3-time unsaturated GPhLs were significantly associated with OS, but only 6 of 19 (32%) saturated GPhLs (P = 0.009, Fisher exact test).

For further analyses, the concentration values of the 43 GPhLs were condensed to one value—a GPhL score—using the Isomap method (a high GPhL score corresponds to high GPhL concentrations). This score correlated significantly negatively with neoadjuvant (n = 10) versus adjuvant chemotherapy (n = 55; P = 0.035), FIGO IV versus FIGO III (P = 0.015; Supplementary Table S1), serum fibrinogen (r = −0.62; P < 0.001), CRP (r = −0.66; P < 0.001; Fig. 1D), and the Eastern Cooperative Oncology Group (ECOG) performance status (1/2/3, n = 8 vs. 0, n = 53; P = 0.009; Supplementary Fig. S7). Positive correlations were revealed between the GPhL score and serum albumin (r = 0.43; P < 0.001), total cholesterol (r = 0.57; P < 0.001), HDL (r = 0.60; P < 0.001), and LDL (r = 0.43; P < 0.001; Fig. 1D). Triglycerides, residual tumor after debulking surgery, grade, body mass index (BMI; Supplementary Fig. S7), age and CA125 at diagnosis, serum gamma-glutamyl transpeptidase (GGT), or alkaline phosphatase were not significantly correlated with the GPhL score (all P > 0.05). The GPhL score, the mean of the significant with OS-associated AAs, and the mean of the nonsignificant GPhLs (but much less pronounced) were significantly correlated with ascites (Supplementary Fig. S8). No correlation was seen with the LDLs in blood, the mean of the nonsignificant AAs, or the acylcarnitines (which showed all the same pattern). Our interpretation is that the tumor mass (i.e., floating tumor cells) present in ascites determined the concentrations of the significant with OS-associated metabolites in blood, but not the concentrations of the other metabolites.

Interestingly, the GPhL score was independent from muscle attenuation and visceral and subcutaneous fat (Supplementary Fig. S7), ascertained by preoperative CT and indicative for tumor cachexia (11). There was also no significant correlation between the GPhL score and specific types of peritoneal tumor spread, military, and non-miliary. The latter is characterized by a more favorable OS, mesenchymal cell characteristics, increased angiogenesis, and adaptive immune system re-activity, but less systemic and local inflammation (6, 12, 13).

To assess the correlation of the GPhL score with the local adaptive immune answer, immune checkpoint markers were assessed by immunohistochemistry staining for CD8 (14), programmed death-ligand 1 (PD-L1), and programmed cell death 1 (PD-1) in 17 ovarian and 16 peritoneal tumors from 20 patients (Supplementary Fig. S9A). Percentages of the following cell types were determined: PD-L1 expressing tumor cells (PD-L1), PD-L1 expressing tumor-infiltrating lymphocytes (PD-L1atTILs), intraepithelial CD8-positive cytotoxic T cells (CD8), and PD-1 expressing intraepithelial CD8-positive T cells (PD-1atCD8). PD-L1 and PD-L1atTILs were significantly positively and independently (from each other and from both other cell types) correlated with the GPhL score (P values, 0.011 and 0.039, respectively; Supplementary Table S2). This means that patients with high GPhL scores showed higher percentages of PD-L1–positive tumor cells and intraepithelial PD-L1–positive TILs (Supplementary Fig. S9B). Univariately, CD8 was significantly positively correlated with the GPhL score (P = 0.030) but no longer in the multiple regression model (P = 0.150).

To assess the interrelations of systemic (serum) and local (ascites) metabolite concentrations, they were measured in ascites from 19 untreated patients, a subset of the 65 patients. Correlation coefficients of GPhLs between serum and ascites were calculated and grouped according to the degree of unsaturation, and again, correlation coefficients increased with degree of unsaturation in both types of GPhLs, O-acyl-O-acyl and O-alkyl-O-acyl (Fig. 2A), but in contrast with the impact on OS (cf.Fig. 1C), even up to the highest degree of unsaturation. To assess the direct impact of tumor mass on the reduction of serum GPhLs and essential AAs, the correlations between serum and ascites were compared in patients with and without residual tumor after surgery (as surrogate marker for the amount of tumor tissue at diagnosis; 8 without and 11 with residual tumor). Interestingly, correlation coefficients for significant OS-associated GPhLs were higher in patients with (median r = 0.79) compared with patients without residual tumor (median r = 0.29). This effect was weaker for nonsignificant GPhLs (median r of 0.61 and 0.33, respectively; Fig. 2B). Similarly, for AAs, only the significant OS-associated AAs were significantly higher correlated in patients with (median r = 0.68) compared with patients without residual tumor (median r = 0.17). In contrast, a minimal negative shift was seen for the 16 not OS-associated AAs (Fig. 2C). In summary, patients with higher tumor load showed higher correlations between serum and ascites GPhL concentrations (more pronounced in significant OS-associated GPhLs) and significant AA concentrations (but not with not OS-associated AAs) compared with patients with lower tumor load. The fact that nonsignificant GPhLs also showed this relation is in accordance with the finding that nonsignificant GPhLs were also associated with OS but below the significance level of 5% FDR (cf. Supplementary Fig. S3). A similar analysis was performed by dividing patients according to their GPhL score (nine highest vs. nine lowest), and again, patients with low GPhLs showed higher correlations between serum and ascites compared with patients with high GPhLs, with a larger difference for significant GPhLs (for details, see Supplementary Fig. S10).

To assess the connection of systemic metabolite concentrations to concentrations in tumor tissues, metabolites from tumor tissues and matched serum samples were compared. Most classes of metabolites, especially the significant OS-associated metabolites (GPhLs, significant AAs, and acetylornithine), correlated positively between serum and tumor tissue, but OS-unassociated AAs and biogenic amines, and sphingolipids (regardless of OS-association) did not (Supplementary Fig. S11A). This means that low metabolite concentrations in serum determine also low concentrations in tumor tissues.

To elucidate which blood fraction(s) or vesicle(s) mostly influence the metabolite measurements in serum, especially the GPhLs, metabolites were measured from different purified serum fractions and vesicles: LDLs, HDLs, vLDLs, exosomes, a vesicle-free soluble fraction, and ultracentrifugation-derived vesicle-free pellets (for isolation of these fractions, cf. Materials and Methods). Interestingly, the serum GPhL-fingerprint was exactly the same as the LDL and vLDL GPhL fingerprint (they cluster very closely in the Isomap), indicating that measured serum GPhLs mainly derive from LDLs and vLDLs. In contrast, serum sphingolipid- and acylcarnitine fingerprints are not identical to any serum fraction's fingerprint. AAs seem to be transported mainly in the vesicle-free soluble fraction of blood, as expected (for details see Supplementary Fig. S11B).

To assess temporal systemic metabolite changes (especially before and after the first therapy usually removing most of the tumor burden), metabolites from follow-up sera were measured and compared with pretreatment concentrations. Follow-up sera of 19 patients revealed clearly that the reduced concentrations of OS-associated GPhLs and AAs are not the result of tumor-independent patient-intrinsic characteristics or permanent changes of metabolism in patients (Supplementary Fig. S12). After debulking surgery and first-line chemotherapy, concentrations recovered to concentrations typical for healthy controls in nearly every patient, even if follow-up sera were taken during recurrence or progressive disease. Only minimal reductions of significant OS-associated GPhLs and AAs were seen in some patients at recurrences, most likely due to the much lower tumor load at recurrence compared with the situation at diagnosis.

To assess the impact of the GPhL score (trichotomized at the 33.3 and 66.7 percentiles) on OS and to examine if this impact is independent from other clinicopathologic and laboratory predictors, univariate and multiple Cox-regression models were estimated. To avoid over-fitting, several parameters (ascites, CA125, CRP, fibrinogen, albumin, HDL, LDL, cholesterol, BMI, and triglycerides) and an AA score (comprised of the five significant OS-associated AAs; highly correlated with the GPhL score, r = 0.78) were each included separately to Cox-regression models, comprised of the trichotomized GPhL score, age, FIGO stage, grade, residual tumor, and the ECOG performance status. Finally, only LDL remained in a Cox-regression model using the Akaike information criterion (AIC) maximization step-down procedure (Table 1A). Age (directly), serum LDL (indirectly), and the trichotomized GPhL score (HR, 0.46; CI₉₅, 0.26–0.81; P = 0.007) were independent predictors for OS in the cohort of 65 HGSOC patients. In Fig. 3A, survival estimates of the multiple Cox-regression model are shown. The trichotomized AA score was univariately significant for OS (HR, 0.54; P = 0.009), but was eliminated from the multiple Cox-regression model by the step-down procedure (presumably due to the high correlation with the GPhL score, which seems to be the more powerful predictor). This means that low concentrations of 43 mainly (poly)unsaturated GPhLs predict unfavorable outcome in HGSOC, independent of typical clinicopathologic factors including ascites, BMI, CA125, classical laboratory measures, and lipoproteins. Four essential AAs (and citrulline) are highly correlated with these GPhLs, but weaker predictors for OS.

rRNA-depleted RNA sequencing data from 32 EpCAM-positive tumor cell samples from primary tumors (n = 7), peritoneal implants (n = 7), and single (n = 10) and aggregated (n = 8) ascites of tumor cells from 16 patients (6) were analyzed for significant correlations with the GPhL score: using a 5% FDR cut-off, 320 genes correlated with the GPhL score, 167 thereof positively and 153 negatively. A signaling pathway impact analysis (SPIA) revealed six pathways significantly associated with 2,054 genes (using a 20% FDR cut-off; Supplementary Table S3): Alzheimer's disease, Parkinson disease, nonalcoholic fatty liver disease, Huntington's disease, alcoholism, and systemic lupus erythematosus. A closer look on these pathways (Supplementary Fig. S13) revealed that they share the same two functional groups of genes: respiratory chain genes (complexes I to V) and histones, which were positively correlated to the GPhL score. Nearly all of the 64 expressed histone genes were positively correlated with the GPhL score (13 of them significantly; only HIST1H2BB and HIST1H3C were nonsignificantly negatively correlated). Reduced histone abundances have been shown to correlate with the number of transcribed genes and also with transcriptional noise, i.e., leaky expression, which is associated with aging in yeast (15). Thus, we correlated the number of expressed genes (coding and noncoding, i.e., number of genes supported by at least one RNA sequencing read-pair) and the number of read-pairs mapped to the genome, but not annotated to any gene from our gene model (putatively representing transcriptional noise) with expressions of histone genes. Indeed, we found a significant negative correlation, as determined by multiple linear regression analyses using the significant histone genes to predict logarithmized numbers of expressed genes and transcriptional noise [especially histone cluster 1, H4j (HIST1H4J), which showed already alone a significant correlation to the number of expressed genes and transcriptional noise, Fig. 4]. This might indicate that lower GPhL concentrations in serum lead to less histone expressions in tumor cells, with subsequently a higher number of different genes expressed and more transcriptional noise (or vice versa!), a many-genes-expressed phenotype.

In addition, the carbohydrate-responsive element-binding protein (ChREBP) gene was negatively correlated with the GPhL score. ChREBP is a main regulator of fatty acid biosynthesis, binding, and activating—in a glucose-dependent manner—carbohydrate response element motifs in the promoters of fatty acid synthesis genes (16).

The gene expression data were also used to compare the GPhL score to published HGSOC molecular subclassification systems, i.e., the Yoshihara subclassification (17), validated in an independent patient cohort (7), the C1-C6 subclassification (18), and the TCGA subclassification (19). The GPhL score was significantly negatively associated with the Yoshihara FIGO III/IV subclass (P = 0.0001), characterized by significantly shorter OS, the TCGA “high-risk” subclass (P = 0.0002), and positively with the C2 (adaptive immune type, P = 0.027) and the C3 (serous low malignant potential and low-grade, P = 0.0009) subclasses, both with favorable survival.

To assess the local and global immune situation in association with the GPhL score, two panels comprised of 56 cyto/chemokines and tumor markers and another panel of 25 cytokines were measured in serum and ascites samples and correlated with the GPhL score. Altogether, ten analytes measured in serum were significantly associated with the GPhL score: Myeloid progenitor inhibitory factor 1 (MPIF1), IL6, osteopontin, CXCL2, soluble epidermal growth factor receptor (sEGFR), leptin, granulocyte-macrophage colony-stimulating factor (GMSCF), IL2R, macrophage inflammatory protein 1-alpha (MIP1a), and -beta (MIP1b); thereof only sEGFR and leptin positively (Supplementary Fig. S14A and S14B). Among the cytokines measured in ascites, osteopontin was significantly negatively, and MIF significantly positively, correlated with the GPhL score (Supplementary Fig. S14C). Osteopontin was significantly associated with the GPhL score in both serum and ascites.

To validate the OS impact of the GPhL score with an independent patient cohort, a robust 100-gene expression signature was developed from RNA sequencing data. This gene signature is comprised of genes which were most significantly positively and negatively associated with the GPhL score. It was applied to an independent validation cohort of 165 serous EOC patients (6, 9), and for each patient, a gene predictor (indicative for the GPhL score) was calculated as follows: median expression of the 50 positively correlated genes minus median expression of the 50 negatively correlated genes. This gene predictor (again trichotomized at the 33.3 and 66.7 percentiles) was used in univariate and multiple Cox-regression analyses (Table 1B). Corrected for the clinicopathologic parameters, age, FIGO stage, grade, and residual tumor (ECOG performance status and laboratory parameters like serum LDL were not available for this cohort), the gene predictor revealed an independent significant impact on OS (HR, 0.65; CI₉₅, 0.50–0.88; P = 0.004), shown as survival estimates in Fig. 3B. In univariate analysis, the gene predictor did not reach significance (HR, 0.83; CI₉₅, 0.63–1.09; P = 0.180).

For the first time, serum metabolites were associated with OS in HGSOC patients, revealing significance in 43 of 89 GPhLs, mainly mono- and polyunsaturated, and 5 AAs, 4 of them proteinogenic and essential for humans. Serum concentrations below typical concentrations in healthy controls were associated with unfavorable survival. A score derived from these 43 GPhLs and an associated 100-gene expression signature, developed from matched RNA-sequencing data, proved the significant impact on OS in test and validation cohorts, independent of age, FIGO stage, grade, residual tumor, ECOG status, serum LDL, other lipoproteins, and laboratory predictors like CRP or CA125. Interestingly, a comparison of the serum GPhL-fingerprint with fingerprints of purified serum fractions revealed clearly that GPhLs in serum correspond to liver-derived LDLs (and the less abundant vLDLs). This is even more interesting as high LDL-cholesterol in serum was an independent positive predictor for OS but the reduced content of mainly unsaturated GPhLs in LDLs a negative one.

The significant OS-associated GPhLs correlated negatively with typical liver-derived systemic inflammation markers such as fibrinogen and CRP and positively with albumin. Further, the GPhLs were associated with ten serum and two ascites cyto/chemokines. Thereof, seven are strongly associated with monocyte/macrophage biology, and all of them showed a negative correlation with GPhLs:osteopontin, which stimulates proinflammatory cytokine production by macrophages (20), the T-cell and monocyte chemoattractant MPIF1, IL6-secreted by macrophages and promoting macrophage differentiation (the pro- or anti-inflammatory effect of IL6 is context dependent; ref. 21), the leukocyte chemoattractant CXCL2—mostly produced by monocytes and macrophages, the macrophage secreted GM-CSF which stimulates the development of macrophages and granulocytes, and MIP1a/b, both attracting macrophages, monocytes, and neutrophils. This is in accordance with the negative correlation of the GPhL score with CRP and fibrinogen and the positive correlation with albumin, all three liver-derived markers for systemic inflammation. Osteopontin, the only cytokine significantly associated with GPhLs measured in both, ascites and serum, is involved in tumor progression and metastasis and is known to be elevated in serum of cancer patients. Moreover, osteopontin is involved in inflammatory processes (20). In summary, GPhLs-low patients showed significantly increased systemic inflammation and an increased monocyte/macrophage activity (innate immune system). The significant correlation of citrulline, a marker for sepsis if reduced, with worse outcome and indirectly with increased systemic inflammation in GPhLs-low patients further supports this interpretation (10).

In contrast, the adaptive immune checkpoint marker PD-L1 on tumor and immune cells was independently positively correlated to GPhL concentrations and also the CD8 tumor infiltration (albeit only univariately), indicating an increased T-cell activity in GPhLs-high patients. This is in accordance with a positive correlation of the GPhL score with the C2 (adaptive immune) subclass and the frequently reported positive prognostic impact of tumor-infiltrating (CD8 positive) lymphocytes in ovarian cancer (14). Further associations of the GPhL score with published subclassification systems of ovarian cancer are in line with our proposed positive impact of GPhLs on OS: negative correlations with two previously described subclasses with unfavorable survival, the Yoshihara FIGO III/IV (7, 17), and the TCGA “high-risk” subclass (19); and a positive correlation with the C3 “benign” subclass (18) indicating a more malignant phenotype for the GPhLs-low (many-genes-expressed) tumors.

Genome-wide gene expression and pathway analyses revealed that low GPhLs were associated with decreased respiratory activity but increased de-novo fatty acid biosynthesis. Both findings are probably adaptations of the tumor cells to the reduced systemic lipid availability which is also reflected in corresponding low GPhL concentrations in tumor tissues. Given the putative functional property of GPhLs as fuel for oxygenic respiration (22) in combination with an unchanged or even increased systemic glucose availability, a direct correlation to the expression of respiratory chain genes indicates this function also in HGSOC. Furthermore, GPhLs, especially (poly)unsaturated ones (desaturation needs oxygen, which is limited in hypoxic cancer tissues and anoxic ascites), as building blocks for increased cell proliferation and growth could explain the negative correlation to the ChREBP gene expression: less available systemic and local GPhLs more de-novo fatty acid biosynthesis.

The most striking result was the direct correlation of GPhLs with histone expressions. Reduced histone abundances have been described in aging (yeast) cells and were shown to be accompanied by a less dense chromatin structure resulting in an increased number of expressed genes and transcriptional noise (15). Maintaining an open chromatin structure at the stem cell stage was already shown to be involved in promiscuous gene expression (23). And indeed, both, the number of expressed genes and transcriptional noise, correlated significantly negatively with histone gene expressions and therefore also with GPhLs. These associations imply that lower GPhL concentrations in serum lead to lower expressions of histones in tumor cells (or vice versa or independently only as nonfunctional correlation), subsequently causing a higher number of different genes expressed and more transcriptional noise. We can only speculate that reduced histone expression is an adaptation to reduced systemic lipid availability or, vice versa, that tumor cells with reduced histone abundances but increased transcriptional activity represent a “neo-pluripotent” phenotype (many-genes-expressed), probably more capable of using extracellular lipids (and essential AAs) as fuel or building blocks, thus depleting them from serum and ascites. A broader transcriptional basis, thus more expressed genes and consequentially more functional pathways, is not only known from aging cells (here probably more a failure of tight gene expression regulation; ref. 15) but also from stem cells, giving these cells a broader biological functionality and development-potency (23, 24), also known as “lineage priming,” first discovered in hematopoietic stem cells but also shown in embryonic stem cells (25, 26). The varying immune system activities related to GPhLs and more innate immunity in GPhLs-low and more adaptive immunity in GPhLs-high patients could also be associated with these different expression profiles. One possible explanation is that in GPhLs-low patients, tumor cells are defined by lower histone expressions leading to total more genes expressed, however presumably not involving the expression of specific (mutated) tumor antigens (i.e., this group of tumors is characterized by global but subtle gene expression changes—many-genes-expressed—and less by single driver-mutations or -pathways). Thus the adaptive immune system is not stimulated by these tumor cells, nonetheless they are recognized, thus triggering (systemic) innate inflammation. In contrast, in GPhLs-high patients, tumor cells are defined by more specifically deregulated gene expressions, presumably expressing more immunogenic tumor antigens recognizable by adaptive immune cells. In addition, these tumors seem to be more dependent on immune escaping mechanisms like the immune checkpoint PD-L1/PD-1-system. The favorable OS of this patient group is probably (partly) due to this increased adaptive immune answer.

If reduced serum GPhLs and essential AAs are a direct effect of tumor metabolism (i.e., if tumor cells consume these metabolites), tumor load in the peritoneal cavity should be positively associated with the correlations of GPhLs and of essential AAs between serum and ascites. And indeed, patients with residual tumor after debulking surgery (used as surrogate marker for tumor load) showed significantly higher metabolite correlations between serum and ascites, especially those which were significantly associated with OS, compared with patients without residual tumor. The same effect was evident for patients with low GPhLs compared with patients with high GPhLs. In addition, OS-associated metabolites correlated with ascites volume (and therefore putatively also with the total amount of tumor cells), but OS-unassociated metabolites and LDLs in blood did not correlate with ascites, further supporting and following Occam's razor (always use the simplest explanation) that the GPhL and essential AA concentration differences in serum and ascites are directly caused by tumor cells. This is further supported by the finding that metabolite concentrations recover to concentrations seen in healthy controls after debulking surgery and first-line chemotherapy. However, there was no correlation of GPhLs with a measure for tumor cachexia (11), nutritional measures (visceral and subcutaneous fat), BMI, or different types of peritoneal tumor spread (miliary vs. non-miliary; refs. 6, 12, 13).

Summarized (Fig. 5, cf. Supplementary Fig. S1), glycerophospholipids in serum, mainly mono- and polyunsaturated ones, together with four essential AAs (and citrulline), correlated between serum, ascites, and tumor tissues, and are strong and independent predictors for OS in HGSOC. Probably, a specific tumor phenotype, associated with low histones leading to a loose chromatin structure and thus high transcriptional activity and diversity, allows tumor cells to deplete these metabolites from serum more efficiently and use them as fuel and building blocks. This specific tumor phenotype seems also to be associated with higher systemic inflammation and a more active innate immunity (especially monocytes/macrophages) but with a lower adaptive immune system reactivity (CD8 TILs and PD-L1/PD-1). Therefore, immune checkpoint inhibition might be less effective in HGSOC compared with other cancer entities, as patients with unfavorable survival are less associated with the PD-L1 immune system inhibition mechanism (27, 28).

Unfortunately, the concrete mechanism(s) how cells take up glycerophospholipids from serum, mainly from lipoproteins like (v)LDLs, is/are not completely understood (29, 30). There are different possibilities: lipoproteins are (i) absorbed or (ii) endocytosed by cells (via e.g., apoB100 binding) or phospholipids are taken up (selectively) by (iii) specific receptors (SR-BI and CD36) or only by (iv) diffusion, driven by concentration gradients between lipoproteins and cell membranes without taking them up. Membrane lipid phosphatidylcholines were shown to be an unexpected source of energy producing triacylglycerols in—at least—hepatocytes (22). Once these mechanisms are defined in more detail, especially in malignant cells, these processes could be used as targets for specific therapies, e.g., by inhibiting uptake of glycerophospholipids from (v)LDLs by tumor cells.

No potential conflicts of interest were disclosed.

Conception and design: A. Bachmayr-Heyda, S. Aust, D. Pils

Development of methodology: A. Bachmayr-Heyda, K. Auer, D. Pils

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A. Bachmayr-Heyda, S. Aust, K. Auer, S.M. Meier, K.G. Schmetterer, S. Dekan, C. Gerner, D. Pils

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A. Bachmayr-Heyda, S. Aust, K.G. Schmetterer, S. Dekan, C. Gerner, D. Pils

Writing, review, and/or revision of the manuscript: A. Bachmayr-Heyda, S. Aust, S.M. Meier, K.G. Schmetterer, C. Gerner, D. Pils

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K. Auer, C. Gerner, D. Pils

Study supervision: S. Aust, D. Pils

Visceral and subcutaneous fat values were analyzed and provided by Thomas Knogler.

This work was supported by funds of the Oesterreichische Nationalbank (Anniversary Fund, project number: 14595). Metabolite measurements were partly funded by the Ludwig Boltzmann Cluster Translational Oncology.

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.