Discussion
To date, carbohydrate antigen 19.9 (CA-19.9) is the only biomarker routinely used, and FDA approved, for the clinical management of PC. However, its primary uses are for prognosis [
26‐
28] and for monitoring high-risk populations [
29]. CA-19.9 has questionable value for average-risk screening [
28,
30], and due to low sensitivity and specificity, the American Society of Clinical Oncology (ASCO) does not recommend the use of CA19.9 for diagnostic screening purposes regardless of symptoms [
31]. In the discovery PC patient population reported herein, only 21 of the 40 patients had CA19.9 levels greater than 35 U/ml (52.5% sensitivity; results not shown), consistent with the abovementioned reports. Accordingly, there are currently no viable means to screen for increased risk of, or early-stage PC.
The non-targeted metabolomics discovery platform used in this study has previously identified early-stage biomarkers mechanistically involved in Alzheimer’s disease [
16], autism [
32], and colorectal cancer [
15]. The key advantages of this platform are that: 1) Samples are processed using a liquid-liquid extraction followed by direct infusion of each extract without chromatography where all molecules are introduced into the system and can therefore potentially be detected. 2) The ultra-high resolution of the FTICR-MS enables mass measurements with accuracy sufficient for the computational determination of elemental composition, and rapid insight into the identities of peaks. 3) Translation of FI-FTICR-MS discoveries into sensitive and cost-effective targeted and quantitative FI-MS/MS assays is seamless due to the high compatibility of the two systems. The high correlation between the non-targeted FI-FTICR-MS and targeted FI-MS/MS results shown in this study validates high-resolution non-targeted metabolomics as a highly sensitive and accurate tool for
de novo biomarker discovery applications.
Overall, the difference in magnitude between the PC patient and control subject serum metabolomes we observed in this study was remarkable. In fact, PCA performed on the entire dataset of 2478 masses, with no prior filtering, clearly discriminated PC patients from controls (not shown). This was due to multiple affected metabolic systems each containing numerous similarly-behaving components. By clustering the FI-FTICR-MS data by Pearson correlation, it was possible to quickly identify these systems. However, the limitation of FI-FTICR-MS, particularly for intact glycerolipids, is that the fatty acid side-chain speciations can only be speculated due to isomerism. For example, differentiating 16:0/18:3 at SN1/SN2 from 16:1/18:2 is not possible. FI-MS/MS therefore represents an ideal complimentary approach because it not only allows for the confirmation of side-chain speciations, but also the investigation of related metabolites that may not have been detected with FI-FTICR-MS due to low abundance, etc.
Using this two-pronged approach, we discovered and confirmed the involvement of three major dysregulated metabolic systems in the serum of PC patients: ultra long-chain fatty acids, numerous choline-containing glycerophospholipids, and vinyl ether-containing ethanolamine phospholipids called plasmalogens. Although most of the individual metabolites alone showed a significant reduction in PC patient serum, the strongest discriminator based on multiple statistical criteria was PC-594 (p=9.9E-14).
The ROC-AUCs based on PC-594 were highly consistent, independent of the platform used or population evaluated in this study (0.98 based on FI-FTICR-MS analysis of the discovery samples, 0.96 based on FI-MS/MS confirmation of the discovery samples, and 0.97 based on the blinded analysis of the US Caucasian patient cohort). On average, the mean PC-594 concentration in PC patients was more than five times lower than control subjects. PC-594 reduction was not observed to correlate with either the magnitude of disease burden (as assessed by stage) or treatment. Interestingly, PC-594 levels were also reduced in IPMN patients. Although IPMN is technically a cystic tumor, it is still a cancer with an invasive component in a high percentage of cases [
33]. These results suggest the possibility that the tumor is not responsible for the reduction, but rather that the reduction precedes the onset of disease, similar to GTA-446 reduction and CRC [
4,
34].
Because the RF predictor based on 11 masses (Figure
4A), showed little improvement in diagnostic accuracy over PC-594 alone (the ROC-AUC based solely on PC-594 was 0.96 versus 0.98 for the RF model), we proceeded with further validation of PC-594 only. A single-analyte assay has several advantages including simpler method development, quantitation, and an easier regulatory approval path. However, our results also showed a strong association between phosphocholine reduction and PC, and therefore evaluating subjects for choline-related deficiency as an independent risk factor should not be disregarded.
Before commenting on the biological implications of the findings, two key limitations of the study should be addressed. First, the current study did not include subjects with non-malignant pancreatic-related conditions such as pancreatitis or jaundice. Since a high percentage of PC patients exhibit jaundice, and up to 5% of subjects with pancreatitis develop PC in a 20-year period [
35], we cannot exclude the possibility that these conditions are also linked to the metabolic effects observed. Although subjects with these conditions would not necessarily be representative of a low-risk target screening population, it will be important to determine whether these conditions affect the biomarkers reported herein, particularly in light of recent reports that jaundice can impact performance of certain PC protein markers [
36].
Second, the sample sizes of the studies were not large. In particular, the clinical diversity of the discovery population was high, including samples from patients collected at time of surgery, following treatment, as well as a low number of cases by disease stage. Although we observed no bias toward any of these variables, (Figures
1B and 6), interpretation (especially for the lack of disease stage effect) should be taken with caution.
The biological implications of reduced systemic levels of long-chain FAs and numerous classes of glycerophospholipids in PC patients are intriguing and warrant further discussion. The long-chain FAs, although only recently reported, represent a large family of 28 to 36 carbon polyhydroxylated and polyunsaturated long-chain fatty acids, originally named gastric-tract acids (GTAs) for their role in CRC [
15]. The prototypical member of the family, GTA-446, has 28 carbons and is reduced in colorectal cancer patients relative to control subjects [
4,
15,
34].
In previous studies, we showed that human serum extracts enriched for selected GTAs protected against inflammation through the down-regulation of NFκB and several pro-inflammatory markers in both human colon cancer and RAW264.3 mouse macrophage cells exposed to lipopolysaccharide [
37]. GTA-treated cells also showed reduced proliferative capacity through a pro-apoptotic mechanism [
37].
In colon cancer, the current hypothesis is that GTAs act analogously to the resolvins and protectins [
38], protecting the body against the accumulation of chronic inflammation over time. Compromised levels with age are suspected to favor the establishment of a pro-inflammatory environment, and ultimately lead to the DNA damage observed in many tumors. PC-594 belongs to the same metabolic system as GTA-446; therefore, it is probable that PC-594, at least to some extent, is also involved in inflammatory processes. Given that PC incidence is low in subjects under age 45 (<3% of cases) and increases with age thereafter (SEER data, 2005–2009), it is tempting to speculate whether the age-related reduction of PC-594 could be causally involved in the establishment of PC.
Given the role of GTAs in inflammation, our current work is focused on determining whether subjects with chronic pancreatitis, an inflammatory condition, have altered levels of PC-594 and other ultra long-chain FAs. Likewise, further investigation of GTA family members across different cancers and inflammatory conditions is warranted for dissecting the specific roles that different isoforms play in the causation of these diseases.
Phosphocholine metabolism has also been previously implicated with PC. For example, results by Yao
et al. showed decreased choline levels in PC tumors via proton MR [
39], and others have shown that human cancer cell growth, including PC cells, can be inhibited by various sphingolipids [
40,
41]. Fang
et al. showed by NMR that rats with PC exhibited lower phosphocholine and glycerophosphocholine compared to rats with chronic pancreatitis [
42]. One of the most convincing
in vivo studies to date functionally implicating choline metabolism to PC was by Longnecker
et al., who showed that rats fed a choline-supplemented diet exhibited significantly reduced PC lesion areas, lesion diameters and numbers of lesions compared to rats fed a choline-devoid diet. This lead the authors to conclude that a choline-deficient diet might have a growth promoting activity [
43].
Choline is also important for pancreatic cell function, as pancreatic acinar cell integrity and the generation of digestive enzymes and insulin secretion are dependent upon high choline phospholipid metabolism [
44]. There is also evidence that reduced sphingomyelin levels may be oncogenic as demonstrated by inhibition of the RAS-MAPK, CyclinD-CDK4/CDK6 and PI3K-AKT axes through the activation of sphingomyelin synthase by a synthetic fatty acid [
45]. Our finding of reduced circulating levels of choline-based metabolites as possible contributing factors to the development of PC is consistent with these observations.
Our observation of reduced PlsEtns (containing the signature vinyl-ether bond at the SN1 position), but not their diacy counterparts, is also intriguing for several reasons. PlsEtns are membrane phospholipids, primarily located in cells of the nervous system and heart [
46,
47], but which are produced exclusively by peroxisomes in the liver [
48]. The vinyl-ether bond at the SN1 position is required for the lipid’s anti-oxidant role and its effect on membrane fluidity, which is relevant to neuronal impairment because it affects vesicular fusion (for review see [
49,
50]). We previously reported that reduced systemic levels of PlsEtns in Alzheimer’s disease patients correlate with levels in the brain and cognitive parameters [
16]. PlsEtns, however, have also been implicated in cancer [
51,
52], and plasmalogen analogues have been shown to exhibit anti-tumor properties [
53]. The consequence of decreased PlsEtn levels, via effects on membrane structure and microdomain architecture, could impact growth factor receptor-mediated signaling.
The findings of this study represent an opportunity for identifying subjects with PC or a high-risk of developing PC. Consider, for example, current CRC screening guidelines, which suggest that the benefit of identifying early-stage CRC (i.e. the increase in 5-year survival) via endoscopic examination in an asymptomatic population with an incidence rate of 0.05% [
54] outweighs the combined risks of complications from endoscopy and late-stage detection mortality. Given the current PC incidence rate of approximately 9.5/100,000 (0.0095%), a blood test with 95% specificity and sensitivity approaching 90% would yield a PC detection rate of approximately 8.6 per 5000 positive tests, or 0.17%. This represents an 18-fold increase in PC risk over average-risk subjects given a positive PC-594 test, and a 3-fold higher incidence rate than the current incidence considered sufficient for colonoscopy-based screening of CRC in the general population. Considering the high mortality rate of late-stage detection, the benefits of endoscopic screening in a small population of high-risk subjects with low PC-594 levels becomes obvious.
Improved clinical benefit was further supported by the results of decision curve analysis, which showed significant net clinical benefit above all p
t
’s greater than 10% compared to screening all subjects with EUS, which was not surprising given the high discriminating ability of PC-594. The results of our studies clearly implicate a circulating reduction of PC-594 in PC, and establish the foundation for designing future prospective trials to determine the net clinical benefit of PC-594 screening in the true average-risk population.
Competing interests
Shawn Ritchie, Wei Jin, Elodie Pastural, Tolulope Sajobe, Dushmanthi Jayasinghe, Bassirou Chitou and Yasuyo Yamazaki were paid employees of Phenomenome Discoveries, Inc. Dayan Goodenowe is a director of Phenomenome Discoveries, Inc. The other authors of have no competing interests. Phenomenome Discoveries, Inc. financed the publication costs of this paper. Part of the results discussed in this paper have been provisionally filed for patent (U.S. application number 13/499,369).
Authors’ contributions
SAR, HA, IT, YY and DBG designed the metabolomic studies. SAR and DBG were the primary authors. IT, HE, HN, MM, YD and MM were responsible for the Japanese trial design, patient enrollment, clinical data management and interpretation of findings. SAR and EP performed multivariate statistics and analysis of FI-FTICR-MS and FI-TQ-MS/MS data. BC and TTS performed the Random Forest, jitter plots, ROC and clinical performance statistical analyses. TTS performed the calibration and decision curve analysis. DJ and WJ developed the FI-TQ-MS/MS assays. TW was responsible for the USA Caucasian study design and blinded analysis. All authors had input, read, and agree with the contents of the manuscript. All authors read and approved the final manuscript.