Abstract

   Background: Pancreatic ductal adenocarcinoma (PDAC) is one of the most
   aggressive human malignancies. The metabolomic approaches are developed
   to discover the novel biomarkers of PDAC.

   Methods: 550 preoperative, postoperative PDAC and normal controls (NCs)
   serums were employed to characterize metabolic alterations in training
   and validation sets by LC-MS.

   Results: The results of PLS-DA analysis indicated that three groups
   could be distinguished clearly and the post-PDAC group is adjacent to a
   normal group as compared with pre-PDAC group. Further results showed
   that histidinyl-lysine significantly increased whereas docosahexaenoic
   acid and LysoPC (14:0) decreased in pre-PDAC patients as compared with
   NCs. And these three markers had a significant tendency to recover
   after tumor resection. The validation set results revealed that for
   CA19-9 negative patients, 92.3% (12/13) of them can be screened using
   these three metabolites. The combination of these markers could
   significantly improve the diagnostic performance for PDAC, with higher
   sensitivity (0.93), specificity (0.92) and AUC (0.97). Moreover,
   network and pathways analyses explored the latent relationship among
   differential metabolites. The glycerolipid metabolism and primary bile
   acid synthesis showed variation in network and pathway analysis.

   Conclusions: These three markers combined with CA199 displayed high
   sensitivity and specificity for detecting PDAC patients from NCs. The
   results indicated that these three metabolites could be regarded as
   potential biomarkers to distinguish PDAC from NCs.

   Keywords: Metabolomics, Pancreatic Ductal Adenocarcinoma, Multivariate
   analysis, Biomarkers

Introduction

   It is estimated that the incidence of pancreatic ductal adenocarcinoma
   (PDAC) will be the second-leading cause of cancer-related deaths by
   2030 due to the intractable detection and the poor prognoses
   [49]^1^,[50]^2. The 5-years survival rate of all PDAC patients has
   remained close to 5% [51]^3^-[52]^5. Clinical symptoms of PDAC patients
   are usually unremarkable in the early stage [53]^6^-[54]^7. More common
   clinical diagnostic methods for PDAC are mainly dependent on imaging
   examination and traditional protein biomarkers [55]^8^-[56]^9. Imaging
   examinations, such as magnetic resonance imaging, computed tomography,
   and endoscopic ultrasonography, have insufficient specificity and
   sensitivity for detecting PDAC in the early stage [57]^10^-[58]^12. On
   the other hand, the best-established serum biomarker is carbohydrate
   antigen 19-9 (CA19-9). Unfortunately, CA19-9 is not only insufficient
   for the early stages but also limited in the sensitivity (59%~64%)
   [59]^13^-[60]^15. So, at the time of diagnosis, only 20% of patients
   can remove their tumors, which could increase 5-years survival rate
   from 5% to 25%. Therefore, there is imperative to identify new
   biomarkers that could help in diagnosis of PDAC, which has become a
   medical emergency [61]^16^-[62]^17.

   As an omics technology, metabolomics enable the global and untargeted
   measurement of small molecular (<1000Da) weight endogenous metabolites.
   It is a useful approach to understand known metabolic pathways and
   biological functional alterations in physiological and pathological
   responses [63]^18. In recent years, metabolomics has been utilized to
   clarify the significant changes of tumor mechanisms and the discovery
   of new diagnostic biomarkers for early diagnosis [64]^19. It is
   essential to distinguish the information between diseased and
   non-diseased status [65]^20^,[66]^21. Ultra-performance liquid
   chromatography-mass spectrometry has the most automated, reproducible
   and high-throughput characteristics, which is the most widely
   analytical platform for metabolomics [67]^22. Thus, more and more
   research groups are taking advantage of metabolomics to the discovery
   of cancer biomarkers and understanding of pathophysiologic processes.

   Numerous studies have reported to seek the multiple circulating
   metabolites signatures to discriminate pancreatic cancer cases from
   non-case controls [68]^23^-[69]^26, and pancreatitis cohorts in some
   researches [70]^27^-[71]^29. These selected individual biomarkers are
   available for the detection of PDAC and the receiver operating
   characteristic (ROC) of them have an outstanding result (AUC>0.9)
   [72]^8^,[73]^11^,[74]^26^,[75]^30^,[76]^31. Although these screening
   modalities are generally able to detect PDAC; none of them have been
   implemented in daily practice so far due to poor consistency of results
   [77]^32. Most of pancreatic cancer metabolomics studies used sample
   size ranging from 40 to 100 while only a few numbers of studies used
   more than 500 samples [78]^33^-[79]^35, and insufficient sample size
   may result in unrepresentative and variable results. Furthermore, there
   shall be a tendency to recover for effective biomarkers after tumor
   resection for the post-operative monitoring. This could be an excellent
   piece of evidence for whether or not it becomes a useful diagnosis
   marker. To date, however, no metabolomics study has investigated the
   relation between the resection of the tumor and the change of
   pancreatic cancer metabolism by comparing preoperative and
   postoperative serum samples [80]^34.

   In the present study, we have performed UPLC/Q-TOF MS based metabolite
   profiling analysis on 550 serum samples to screen out critical
   metabolite alterations that may discriminative biomolecules for PDAC
   diagnosis through utilizing preoperative and postoperative pathology in
   training cohorts. Combined with clinical information of the PDAC
   patients, three discriminative metabolites (Docosahexaenoic acid,
   LysoPC (14:0) and Histidinyl-Lysine) were determined to be independent
   predictors for PDAC diagnosis and its diagnostic performance was
   confirmed via independent validation analysis. The performance of three
   discriminative metabolites in PDAC was evaluated, and they provided a
   highly accurate classifier for delineating PDAC patients from NC with
   >97% accuracy (AUC = 0.97). Levels of three postoperative
   discriminative metabolites were closed to normal controls compared with
   paired preoperative PDAC group. In addition, correlation network and
   pathway analysis were carried out to understand the inter-relationship
   among discrepant metabolites. These results demonstrated the potential
   capability of the three metabolic biomarkers could be utilized to
   distinguish PDAC from NCs.

Materials and Methods

Sample Collection

   550 serum samples from 431 populations were involved in this study
   including preoperative (pre-PDAC) and postoperative (post-PDAC)
   patients with PDAC and normal controls (NC). The training set consisted
   of pre-PDAC patients (n=185) and normal controls (n=146). Of the 185
   pre-PDAC patients, 87 pairs of postoperative samples were collected.
   The validation set included another new PDAC preoperative samples
   (n=50), pairs of postoperative samples (n=32) and normal controls
   (n=50). All patients were diagnosed with pancreatic cancer for the
   first time and had no treatment before sampling, and were recruited and
   pathologically confirmed from the Affiliated Tumor Hospital at the
   Harbin Medical University; the serum samples from the healthy
   volunteers were obtained from the Fourth Affiliated Hospital of Harbin
   Medical University. Informed consents were obtained from all the
   enrolled participants before taking part in this study. The malignant
   severity was assessed by using the TNM classification system (the 8th
   edition of AJCC) and differentiation degree. The preoperative serum
   samples were obtained in the next morning after the patients were
   hospitalized and the postoperative ones were sampled in the morning on
   the seventh day after the operation when the pancreatic function was
   recovered. Control subjects were recruited on the basis that they had
   no history of cancer and the serum levels of the tumor markers CEA,
   CA19-9 and AFP were negative. Clinical information of all subjects was
   shown in Table [81]1. Notably, nearly half of the samples came from
   samples after surgery, leading to a high percentage of patients in
   stage I. And the proportion of patients with unknown staging
   information exceeds over 40% due to the mission of staging information
   for patients without surgery.

Table 1.

   Pathological and clinical characteristics of subjects in training set
   and validation set
   Characteristics Training set Validation set
   NC (n=146) PDAC (n=185) post-PDAC (n = 87) NC (n=50) PDAC (n=50)
   post-PDAC (n = 32)
   Age ≤40 17 24 2 1 3 3
   41~50 41 29 18 5 5 10
   51~60 52 63 30 19 19 8
   ≥61 36 69 37 25 25 11
   Sex male 88 109 52 25 25 11
   female 58 76 35 25 25 21
   Diabetes yes 0 14 7 0 5 2
   no 146 130 53 50 37 28
   unknown 0 41 27 0 8 2
   Hepatitis B yes 0 7 4 0 2 0
   no 146 138 65 50 33 21
   unknown 0 40 18 0 15 11
   TNM stage I — 85 69 — 28 17
   II~IV — 28 7 — 14 9
   unknown — 72 11 — 8 6
   Differentiation poor — 55 28 — 11 10
   moderate — 42 26 — 11 4
   high — 41 30 — 19 13
   unknown — 47 4 — 8 5
   Jaundice yes — 65 32 — 12 7
   no — 85 38 — 25 16
   unknown — 35 17 — 13 9
   Surgical method 1 — — 22 — — 14
   approach method 2 — — 40 — — 8
   unknown — — 25 — — 10
   CA19-9 ≤37 — 34 22 — 13 8
   >37 — 95 58 — 21 13
   unknown — 56 7 — 16 11
   [82]Open in a new tab

   Method 1: pancreatectomy; Method 2 pancreaticoduodenectomy.

Reagents and Chemicals

   HPLC grade acetonitrile and methanol were purchased from Fisher
   Scientific (Waltham, MA, USA); formic acid (HPLC grade) was produced by
   Fluka (St. Louis, MO, USA); deionized water was provided by a Milli-Q
   ultrapure water system (Millipore, Billerica, USA).

Sample Preparation

   To provide a measurement of the stability and performance of the
   system, quality control samples (QCs) were prepared by pooling equal
   volume of supernatant of all samples in the identical corresponding
   dataset (i.e., training set or validation set). All serum supernatant
   were carefully collected with non-anticoagulant vacuum tubes and
   immediately centrifuged at 4000× g for 10 min at room temperature. The
   sample preparation was done according to the method described in our
   previous report [83]^36. Serum samples were thawed on ice and 100 µl
   aliquots were mixed in 300 µl pre-cooled methanol/acetonitrile (1:1)
   for protein precipitation. Finally, the dried residue obtained after
   freeze-dried was re-dissolved by 100 µl of 50 % methanol.

UPLC-Q/TOF MS Analysis

   To ensure stability during analysis, samples were analyzed for quality
   control at the beginning and at the end of each running batch.
   According to the method described in a previous report, the injected
   sample volume 5 µl. Chromatographic separation was performed by the
   ultra-performance liquid chromatography (UPLC) system (Waters, Milford,
   USA) using a Waters BEH C18 column (2.1 mm × 100 mm, 1.7 µm) (Waters,
   Milford, MA) kept at 40 °C in ESI (+) and ESI (-). The elution flow
   rate was 0.30 mL/min to avoid insufficient nebulization. The optimized
   elution gradient was performed as follows: 0-0.5 min 1 % eluent A ,
   0.5-3.5 min 1-53 % eluent A , 3.5-7.5 min 53-70 % eluent A, 7.5-9 min
   70-90 % eluent A , and then maintained at 9-13 min 90 % eluent A
   followed by alternating the gradient back to 13.1 to 15 min 1 % eluent
   A (0.1 % formic acid - acetonitrile (A) and 0.1 % formic acid - water
   (B)).

   MS data identification and MS/MS acquisition were both performed in a
   dual electrospray ion source (Agilent, Santa Clara, CA, USA) with a
   6520 series accurate quadrupole time-of-flight mass spectrometer (Q-TOF
   MS). MS data was collected in the positive and negative mode equipped
   with a scan rate of 1.5 spectra/s and the mass range was from 50 to
   1100 m/z. The parameters for the acquisition were using the following
   settings: the capillary voltage, 4 and 3.5 kV in the positive and
   negative mode, respectively; the gas temperature, 330 °C; the flow
   rate, 10 L/min; the fragmentor, 100V; and the skimmer, 65 V.

Statistical Analysis

   Our initial analysis of the whole group was to process mass spectra
   data obtained by LC-MS using Qualitative Analysis B.04.00 to extract
   and align peaks. In the training set study, principle component
   analysis (PCA), partial least square discriminant analysis (PLS-DA) and
   orthogonal projection to latent structures discriminant analysis
   (OPLS-DA) were conducted to demonstrate that metabolomic profiling of
   PDAC patients and NC. 100 permutation tests of cross-validation were
   used to avoid over-fitting and to certify the credibility and stability
   of the PLS-DA and OPLS-DA models [84]^37. Furthermore, the potential
   differential metabolites were selected via a univariate nonparametric
   Kruskal-Wallis rank sum test (p<0.05) and a multivariate random forest
   (RF) model (VIMP>1) was designed [85]^38. To exclude the influence of
   gender, age and jaundice, the potential metabolites were selected
   through employing a multivariate logistic regression considering the
   three confounding factors. And the potential metabolites which were
   significant call-back in the paired postoperative serum samples were
   consequently used for the subsequent analysis. To obtain the
   discriminative metabolites with satisfactory predictive performances
   combined with CA19-9, we followed the criteria for discriminative
   metabolite determination from differential metabolites: (1) displaying
   a Pearson correlation coefficient (CC) with CA19-9 smaller than 0.15;
   (2) displaying a univariate AUC larger than 0.8 in the discrimination
   between PDAC and NC.

   Hierarchical cluster analysis (HCA) was performed to visualize the
   significant intensity differences in the concentration levels of these
   differential metabolites in a heatmap. Subsequently, Pearson
   correlation analysis and multivariate logistic regression were applied
   to evaluate whether the differential metabolites were correlated with
   CA19-9 and the independent clinical factors and the PDAC diagnosis,
   respectively. Receiver operating characteristic (ROC) analysis was used
   to calculate the area under the ROC curve (AUC), sensitivity, and
   specificity values for the model to evaluate the predictive power of
   the discriminative metabolites alone and together with CA19-9 for PDAC
   diagnosis performance. The optimal cut-off value of the model was
   determined from its ROC curve. In the validation set study, the
   diagnostic model was evaluated using the AUC, sensitivity, specificity,
   and accuracy values observed at the cut-off value obtained in the
   training set study [86]^39. Correlation network and a pathway analysis
   were performed to further illustrate the latent relationship between
   the differential metabolites in PDAC.

Metabolite Identification and Screening

   Identification of metabolites was completed as described in our
   previous work. Shortly, accurate mass measurements were subject to
   database searches in the public databases METLIN
   ([87]http://metlin.scripps.edu/index.php). According to the RT, m/z and
   MS/MS spectrum of differential metabolites, they were well matched with
   those of authentic standards or confirmed spectrums in the public
   databases HMDB ([88]http://www.hmdb.ca/), METLIN, as well as MassBank
   ([89]http://www.massbank.jp/) [90]^40.

Results

Metabolic Profiling of PDAC and NC

   Metabolic profiling of PDAC and NC of 550 serum samples of the training
   set and validation set were acquired using UPLC/Q-TOF MS. The workflow
   for the metabolomics data analysis was presented in Figure [91]1. In
   training cohort, the typical basic peak chromatograms (BPC) of the
   pre-PDAC (n=185), post-PDAC (n=87) and NC (n=146) group in both the
   positive and negative ionization mode were shown in [92]Figure S1.
   There was remarkable fluctuation of the height at the same arriving
   time of the chromatographic peaks of NC, pre-PDAC and post-PDAC group
   whether in positive and negative ionization mode. The PCA score plot
   for subjects in the training set was shown in [93]Figure S2A,
   displaying a visible separation in the scores plots from patients with
   malignant disease and normal controls. Furthermore, the PLS-DA score
   plot (model parameters R2 = 0.88, Q2 = 0.87; Figure [94]2A) showed a
   clear separation of PDAC patients from NCs and no obvious over-fitting
   was observed in the permutation test (Figure [95]2B). These analyses
   indicated that there were obvious differences in the serum metabolic
   profiles between the PDAC and NC. Further, the OPLS-DA score plot
   described the differences of metabolic profiling in paired samples
   before and after surgery (Figure [96]2C). Then, the OPLS-DA score plot
   showed a clear separation of pre-PDAC patients from post-PDAC patients
   and no obvious over-fitting was observed in the permutation test
   (Figure [97]2D).

Figure 1.

   [98]Figure 1
   [99]Open in a new tab

   The workflow of the metabolomics data analysis.

Figure 2.

   [100]Figure 2
   [101]Open in a new tab

   Metabolic profiling analysis among NC, pre-PDAC and post-PDAC groups.
   The score plot for PLS-DA (A) to discriminate pre-PDAC (n=185) and
   NC(n=146); and cross-validation plot obtained from 100 permutation
   tests (B); The score plot for OPLS-DA (C) to discriminate pair-wise
   pre-PDAC (n=87) and post-PDAC (n=87); and cross-validation plot
   obtained from 100 permutation tests (D); Three-dimensional score plot
   for PLS-DA (E) to discriminate pre-PDAC (n=87), post-PDAC (n=87) and NC
   (n=146); and cross-validation plot obtained from 100 permutation tests
   (F).

   To seek for the metabolic changes caused by PDAC, we studied the
   effects of surgical resection on metabolomic profiles. The PCA
   ([102]Figure S2B), three-dimensional PLS-DA (Figure [103]2E) and
   OPLS-DA ([104]Figure S2C) pattern recognition techniques were applied
   to analyze the NC, pairing pre-PDAC and post-PDAC serum. The
   three-dimensional PLS-DA and OPLS-DA score plot for the subjects in the
   training set indicated that call-back postoperative samples could be
   separated from preoperative samples and normal samples (model
   parameters R2 = 0.68, Q2 = 0.64; R2=0.42, Q2=0.36), respectively.
   Notably, pre-PDCA patients were far away from the negative controls and
   postoperative patients, whereas the post-PDAC patients were located
   closely to NCs. It was evident that the good separation performance was
   achieved in PLS-DA and OPLS-DA model and the results of cross
   validation were reliable (Figure [105]2F and [106]Figure S2D). These
   results revealed that the postoperative metabolic profiles had a
   tendency to recover after tumor resection.

   Considering the influence of jaundice on the systemic metabolism, the
   PLS-DA model metabolic profiling analysis on NC, PDAC with and without
   jaundice groups were performed. The result showed that PDCA patients
   were far away from the negative controls. Notably, there was a certain
   tendency to separate PDAC with and without jaundice group ([107]Figure
   S3). This result suggested that the jaundice might influence the
   selection of differential metabolites, and jaundice should be added as
   a confounding factor in screening differential metabolites.

Selection and Identification of Differential Metabolites

   On the basis of the metabolic profiling, pairwise comparisons of groups
   were carried out to further explore the differential metabolites
   responsible for the differences between pre-PDAC and NC. There were
   8757 ions were found by LC-MS, from which there were 116 ions were
   selected as differential metabolites by Kruskal-Wallis rank sum test
   (p<0.05) and multivariate random forest (VIMP>1), which have been
   identified by MS-MS. In addition, to exclude the disordered metabolites
   caused by jaundice, gender, and age, the logistic regression was used
   to analyze the 116 differential metabolites. The results showed that 11
   potential differential metabolites affected by jaundice, gender and age
   have been removed ([108]Table S1), and it is proved that three of them
   has a relationship with jaundice or bilirubin by previous studies
   [109]^41^-[110]^43. Finally, the significant call-back metabolites were
   screened comparing the preoperative and postoperative PDAC pairs of
   samples, which were regarded as useful markers. 31 metabolites were
   found have significant call-back (the adjusted t-test's p value <0.05),
   including 17 negative ions mode (ESI-) and 14 positive ion mode (ESI+)
   ([111]Table S2).

   To provide a different perspective into the group segregation, the
   HCA-heatmap for all the differential metabolites was presented in
   Figure [112]3A. In the HCA-heatmap diagram, the pre-PDAC group
   observations were completely separated from the post-PDAC and NC group.
   Overall, red cluster represented masses with mean elevation of 16
   metabolites among PDAC patients, while green clusters represented
   masses with reduction of 15 metabolites. Additionally, the post-PDAC
   subjects were generally similar to NCs. This method was the same as the
   grouping patterns shown in the PLS-DA score plot.

Figure 3.

   [113]Figure 3
   [114]Open in a new tab

   HCA-heatmap and the expression of discriminative metabolites. (A)
   HCA-heatmap plot indicating relative levels of differential metabolites
   in samples of the training set. (B) Box plots for comparing
   concentration levels of the three discriminative metabolites in
   different groups in the training set. (C) Box plots for comparing
   concentration levels of the three discriminative metabolites in
   different groups in validation set. * p< 0.05 ** p<0.01 *** p< 0.001
   **** p< 0.0001.

Diagnostic Performance and Verification of Discriminative Metabolites in
External Validation Set

   Three of these difference metabolites, docosahexaenoic acid (FA_1),
   LysoPC (14:0) (LysoPC_1), histidinyl-Lysine (DP_1) have been selected
   by a series of analysis processes (|CC|<0.15 and ACU > 0.8), and might
   be useful for PDAC diagnosis and prognosis. The external validation
   set, another batch of serum sample including NC (n=50), pre-PDAC (n=50)
   and post-PDAC cases (n=32), was collected and analyzed to validate the
   reliability of these three potential marker candidates. The same
   methods of sample pretreatment, instrumental detection, and data
   analysis were utilized. These three metabolites, docosahexaenoic acid,
   LysoPC (14:0) and histidinyl-Lysine showed significant differences (p <
   0.05), and similar variable tendencies with those of the training set
   (Figure [115]3B and Figure [116]3C). Since the surgical approach
   (distal pancreatectomy and pancreaticoduodenectomy) has a totally
   different postoperative recovery, we examined the influence of surgical
   approach on three discriminative metabolites ([117]Figure S4). The
   results showed that three discriminative metabolites were unaffected by
   surgical approach of patients.

   ROC curves of Docosahexaenoic acid, LysoPC (14:0) and Histidinyl-Lysine
   are shown in the validation set, respectively (Figure [118]4A). The
   sensitivity (Se), specificity (Sp), and area under the ROC curve were
   showed in [119]Table S3 for three discriminative metabolites.
   Furthermore, three metabolites were defined as a combinational marker
   with favorable classification capability. The Se, Sp and AUC were 0.93,
   0.92 and 0.97 for using this combinational marker to distinguish PDAC
   from NC. Further, the combinational markers with CA19-9 in the
   prediction of PDAC displayed a Se, Sp and AUC of 0.95, 0.98 and 0.99,
   respectively ([120]Table S3). These AUC values indicated a satisfactory
   performance in the validation data sets, with remarkable sensitivity
   and specificity to accurately stratify subjects into correct groups. In
   addition, the dynamic changes in the normal controls and PDAC patients
   in different pathological stages (TNM staging system) were investigated
   ([121]Figure S5A). The results showed that the level of
   Histidinyl-Lysine was related with progression of PDAC. Besides,
   Docosahexaenoic acid, LysoPC (14:0) and Histidinyl-Lysine showed an
   excellent performance (AUC>0.8) for stage I PDAC patients in training
   set ([122]Figure S5B). These findings suggested that the discriminative
   metabolites might be useful for PDAC early diagnosis and prognosis.

Figure 4.

   [123]Figure 4
   [124]Open in a new tab

   Evaluation of potential biomarkers. Classiers are the biomarker
   signature generated in the training set and presented here for the
   validation set. (A) Scatter plot for graphical representation of the
   biomarker signature score. Y-axis score of biomarker signature with the
   cut-off ≥ 0.41 and CA19-9 on the X axis with the cut-off ≥ 37 U/mL (<
   37µg/mL as CA19-9-negative). (B) Blue circles are pancreatic cancer
   (n=34).

   Moreover, for the CA19-9-negative patients from the validation set, the
   combinational markers had a more ideal accuracy (Figure [125]4B). It
   was noteworthy that the CA19-9 value some of these cases were < 37
   µg/mL, thus the CA19-9-negative patients cannot be distinguished by the
   serum CA19-9 level. 92.3% PDAC patients (12/13) who could not be
   screened by CA19-9 showed a positive result through the three candidate
   diagnostic metabolites. For one third (12/34) of the patients, our
   results would help to improve the diagnostic workup and treatment
   stratification. These results indicated that the discriminative
   metabolites could provide a comparable diagnostic performance of CA19-9
   and the prediction of CA19-9-negative patients, which allow these
   metabolites potentially contribution to PDAC diagnosis in clinical
   practice.

   Notably, after tumor resection, the postoperative serum level of LysoPC
   (14:0) inclined to normal level. Similarly, it was clear that most of
   the LysoPCs (LysoPC (15:0), LysoPC (P-16:0), LysoPC (17:0) and LysoPC
   (20:4(8Z,11Z,14Z,17Z)) were down-regulated in the preoperative
   conditions in comparison to healthy controls, and the elevation was
   obvious after resection of tumor in the training and validation set
   ([126]Figure S6A and [127]Figure S6B). Compared with pre-PDAC serum
   patients, post-PDAC patients have a tendency to return to normal. Due
   to tumor removal, the change of body metabolism made the level of
   LysoPC (14:0) and LysoPCs family call back.

Correlation Network Analysis

   A correlation network was built on the basis of exploiting the latent
   relationships between the differential metabolites in PDAC, which
   ensured the robustness and reliability of the network construction. A
   total of 17 nodes and 26 edges were recruited in the network diagram in
   a circular layout on the criteria of a correlation coefficient ≥0.6
   (Figure [128]5A). In accordance with the molecular composition and
   transforming relationship of metabolites of different classes, the
   entire network could be generally divided into two subnetworks.
   Glycerophospholipids (LysoPCs and LysoPEs) showed down-regulated
   concentration levels (blue nodes) in PDAC patients whereas steroids
   (ST) and bile acids (BA) were up-regulated (red nodes). The
   intra-category gathering landscapes could be clearly observed in
   metabolites of different classes in the network diagram, suggesting the
   underlying transformation of substances and energy in PDAC.

Figure 5.

   [129]Figure 5
   [130]Open in a new tab

   Analyzed of the correlation network and pathways altered in PDAC. (A)
   Correlation network constructed with 18 differential metabolites
   (Pearson correlation analysis, |r| > 0.6). Blue sub-network constructed
   with glycerophospholipids (LysoPCs and LysoPEs). Red sub-network
   constructed with steroids (ST) and bile acid (BA). Nodes in red and
   blue represent the metabolites down-regulated and up-regulated in PDAC,
   respectively. (B) Significantly changed pathways. Disordered pathways
   in PDAC group; small p value and big pathway impact factor indicate
   that the pathway is greatly influenced.

Pathway Analysis

   To further investigate the biochemical perturbation correlated with
   PDAC, an overview of the systematic metabolome changes on the basis of
   pathway analysis were conducted. PDCA-induced metabolic perturbation
   was analyzed from the perspective of pathway enrichment analysis
   combined with the topology analysis. The biological pathways involved
   in the metabolism of these 31 metabolites and their biological roles
   were determined by enrichment analysis using MetaboAnalyst (Figure
   [131]5B). A total of 10 matched metabolic pathways ([132]Table S4) were
   shown according to p values from the pathway enrichment analysis
   (y-axis) and pathway impact values from pathway topology analysis
   (x-axis), the most impacted pathways colored in red. More attention
   should be paid to pathway with high impact values and pathway
   enrichment analysis (p<0.05). The perturbation of metabolism pathway
   includes phenylalanine, tyrosine and tryptophan biosynthesis,
   ubiquinone and other terpenoid-quinone biosynthesis and so on. These
   metabolic anomalies were found to be primarily involved in amino acid
   metabolism, lipid metabolism, and energy metabolism.

Discussion

   PDAC is burdened with a 5-year survival rate of around 5% and will be
   the second leading cause of cancer-related death by 2030 [133]^2.
   Therefore, it is necessary to improve the screening and diagnostic
   method for PDAC. Metabolomics, the 'omics technique' subject to
   environmental influences, has been proposed to be useful for
   identifying new biomarkers for PDAC early diagnosis [134]^44. PDAC has
   a significant heterogeneity within the tumors of individuals
   [135]^45^,[136]^46. This calls for large sample sizes to ensure
   adequate representation of subtypes [137]^47. Furthermore, biomarker
   development programmers required samples to be separated into
   independent training and validation sets [138]^47. In our metabolomics
   analysis, this conventional route that benefits from a large population
   [139]^46 was adopted in this study. And the samples were divided into
   training and validation sets, preoperative and postoperative, which
   guaranteed the reliability of the results. In order to select
   discriminative metabolites, we employed the strategy that was
   considered to be the influence of clinical factors (gender and age),
   which enhances the clinical reliability for epidemiological studies.
   Therefore, our metabolomics approach is acceptable as a screening
   method for large populations.

   To screen out the discriminative metabolites that have satisfactory
   predictive performances alone or combined with CA19-9 from differential
   metabolites, the three criteria were followed in training set: (a)
   employing a Pearson correlation analysis to exclude the metabolites
   that have the correlation coefficient with CA19-9 greater than 0.15;
   (b) employing an area under the curve (AUC) value larger than 0.8. One
   important aspect of the data-modeling procedures lays in the predictive
   ability in terms of sensitivity (Se), specificity (Sp), and area under
   the ROC curve (AUC) in the external validation set distinguishing
   malignant pancreatic disease from normal controls. Previously
   metabolomics efforts have been made to compare PDAC and control
   samples. However, it is difficult to apply to the clinic because their
   models consist of many different metabolites and AUC of their model
   only maintained in 0.7~0.8 in external validation set [140]^52^,
   [141]^53. Jiang and colleagues suggested TSGF as a candidate serum
   biomarker for pancreatic cancer and found that it displayed 91.6%
   sensitivity and 83% specificity [142]^54. However, its sensitivity for
   early stages pancreatic cancer was decreased to 60.0~75.0%. In our
   study, to validate the reliability of these three potential marker
   candidates, we collected and analyzed another batch of serum sample
   including NC cases (n=50) and PDAC cases (n=50). The same methods of
   sample pretreatment, instrumental detection, and data analysis were
   utilized. Our diagnostic performance sensitivity (0.93), specificity
   (0.92) and AUC (0.97) of combinational marker were much more enhanced
   than CA19-9. Combinational markers performed an accuracy of 92.3% for
   CA19-9 negative patients (12/13), which provide a complement to the
   analysis for unsatisfactory performance of CA19-9. The combinational
   markers with CA19-9 in the prediction of PDAC displayed a Se, Sp and
   AUC of 0.95, 0.98 and 0.99, respectively. Indeed, our combinational
   markers effectively assist the diagnostic performance of CA19-9.

   For the discriminative metabolites, docosahexaenoic acid and LysoPC
   (14:0) were down-regulated, while histidinyl-lysine was up-regulated in
   PDAC patients. As the complex structure of LysoPCs, UPLC-MS is the best
   way to determine accurately the levels of each individual LysoPCs from
   minimal amounts of serum. LysoPCs are a class of chemical compounds
   that are derived from PC [143]^48. In our study, not only LysoPC (14:0)
   but also the other members of the LysoPCs family (LysoPC (15:0), LysoPC
   (P-16:0), LysoPC (17:0) and LysoPC (20:4)) were down-regulated in PDAC.
   It is revealed that the LysoPCs might relate with the carcinogenesis
   and progression of PDAC. The long chain dietary polyunsaturated fatty
   acid have been found to enhance various cellular responses that reduce
   cancer cell viability and decrease proliferation both in vitro and in
   vivo [144]^49^-[145]^51. A decrease in docosahexaenoic acid indicates a
   disorder of fatty acids. In addition, a dipeptide is an organic
   compound derived from two amino acids which can identical different.
   Although dipeptides were generally considered as incomplete breakdown
   products of protein digestion or protein catabolism, the specific
   metabolic mechanism of dipeptides in PDAC patients' remains rarely
   reported. It indicated that the amino acid metabolism was disordered
   and may bring disturbance of body metabolism in PDAC.

   Due to alterations in the tumor cell and systemic metabolism, PDAC
   causes changes in circulating metabolites, which is central to the
   biology of PDAC [146]^55. To capture the differential metabolites
   relationships in global changes, network analyses have been widely
   applied in metabolomics studies [147]^56. In this study, phenylalanine,
   tyrosine and tryptophan biosynthesis, ubiquinone and other
   terpenoid-quinone biosynthesis were the majority of perturbed metabolic
   pathways.

   In conclusion, our study showed that the discriminative metabolite
   selection strategy can readily and effectively be applied to serum
   metabolomics on the basis of a multivariate analysis. The selected
   diagnostic metabolites not only have the ability to diagnose PDAC from
   NCs, but also can effectively improve the diagnostic performance of
   CA19-9. All of diagnostic metabolites had a tendency to recur in
   postoperative samples, which suggests that the perturbation is coming
   from the tumors. Moreover, the correlation network and pathway analysis
   presented the relationships between discriminative metabolites and the
   disturbed biological mechanism in PDAC's development. These results
   will not only provide the potential for the improvement in diagnostic
   accuracy, but also the identification of altered metabolic pathways
   between PDAC and NCs, which may help us to understand the mechanisms of
   PDAC.

Supplementary Material

   Supplementary figures and tables.
   [148]Click here for additional data file.^ (2.1MB, pdf)

Acknowledgments