The present study was reviewed and approved by the Tianjin University of Traditional Chinese Medicine. The samples used in the present study originated from the same volunteers used by Fang et al (9), and so shared demographic and clinical characteristics. The group included 36 patients with clinically diagnosed PD and 27 healthy controls from the Affiliated Hospital of Tianjin University of Chinese Medicine (Tianjin, China) and Tianjin Maternity Hospital (Tianjin, China). Urine was collected from PD patients and healthy controls during the luteal regression stage, and informed consent was obtained from all participants prior to sample collection. Each volunteer provided written answers detailing age, weight, height, personal and family history of menstrual cramps, and dysmenorrhea pain integral. The pain integral was >8 in all patients with PD. The selection criteria for patients with PD were in accordance with the diagnostic criteria established by the People's Republic of China Ministry of Health Pharmaceutical Council (10) and the National Higher School Teaching Materials Obstetrics and Gynecology, Seventh Edition (11).

Urine samples were collected as described previously (9), and collection was consistent with the clinical inclusion criteria (morning urine collected three days prior to menstruation, and the volume of each sample recorded). Following this, each sample was loaded into 10 ml centrifuge tube, and 1% sodium azide was added. The samples were subjected to centrifugation at 765 × g at 4°C for 15 min. The supernatant was stored at −80°C until analysis.

Urine samples and quality control (QC) samples were prepared as previously described (9). QC samples were mixture of urine from PD patients and healthy controls. Urine samples were prepared as follows: Samples were thawed at room temperature and centrifuged at 8,497 × g at 4°C for 10 min. Next, 200 µl liquid supernatant was added to 200 µl pure water. The liquid was subsequently vortexed for 1 min and centrifuged at 14,360 × g at 4°C for 15 min. Finally, all samples were injected into UPLC-Q/TOF mass spectrometer for analysis. QC samples were prepared as follows: 8 samples were selected from each group by random sampling. QC samples were centrifuged at 8,497 × g at 4°C for 10 min. Liquid supernatant (200 µl) was then added to 200 µl pure water, vortexed for 1 min and centrifuged at 14,360 × g at 4°C for 15 min. QC samples were analyzed once every 6 h and used to test instrument stability, ensuring that conditions remained the same throughout the analysis.

The liquid chromatograph used was the Waters ACQUITY UPLC system (Waters Corporation, Milford, MA, USA). Supernatant (5 µl) was injected into ACQUITY UPLC BEH C18 columns (2.1 mm × 100 mm; 1.7 µm; Waters Corporation). The column temperature was set at 45°C and the flow rate was 0.3 ml min-1. The gradient system consisted of mobile phase A (0.1% formic acid in water) and mobile phase B (0.1% formic acid in acetonitrile) as follows: 0–8.5 min, 1–25% B; 8.5-11 min, 25–50% B; 11–13 min, 50–90% B; 13–15 min, 90–99% B; 15–17 min, 99% B; 17–18.5 min, 99–1% B; 18.5-20 min, 1% B (9).

Mass spectrometry was performed on a Waters Micromass QTOF Micro Synapt high definition mass spectrometer (Waters Corporation). Electrospray ionization (ESI) was used in positive mode. Ion source parameters were as follows: Capillary voltage, 3.0 kV; cone voltage, 30 V; nebulizer pressure, 350 psi; nitrogen gas temperature, 325°C; cone gas flow, 50 l/h; desolvation gas flow, 600 l/h; source temperature, 120°C; desolvation temperature, 350°C; 0.1 sec (interval 0.02 sec) collected once spectrum data and scanned at mass range of m/z 50–1,100.

Data was initially exported using MarkerLynx V4.1 (Waters Corporation) with peak discovery, peak alignment and raw data filtering to determine potential discriminant variables. A data array that included retention time, m/z values and normalized peak area was then obtained. The multivariate data matrix was exported into SIMCA-P 12.0 software (MKS Instruments, Andover, MA, USA) for PCA and PLS-DA analysis. The PCA model is a statistical sample of the principal contradiction reaction and can resolve the main factor multivariate data, as well as reflect the main features. The data space is compressed, and characteristics of multivariate data in a low-dimensional space are expressed through visual effects. The PLS-DA model was fitted following the adaptation of the data. According to the classification pattern recognition model, compounds were chosen depending on whether they have an important contribution. These data further validate the differences in the compounds between PD patients and healthy controls through metabolic differences in the luteal regression stage (12,13). Cross-validation tests were sorted through for verification. Evaluating urine PLS-DA models requires a higher R2Y value, and Q2 values for these parameters typically. R2X, R2Y and Q2 are larger in the present model, and the R2Y and Q2 are closer to 1, demonstrating the accuracy of the model result (14). Variable-importance plots (VIP) values >1 were used to screen potential biomarkers. Unpaired Student's t-tests were used to determine statistically significant differences between biomarkers in patients with PD and healthy controls. P<0.05 was considered to indicate a statistically significant difference.

Resulting candidate biomarkers were selected based on their molecular weights and use of the formula to predict the elemental composition of a compound. The database was then used to retrieve documents and metabolite candidates, and the candidates were finally confirmed using metabolic standards, references and tandem mass spectrometry information, such as the Human Metabolome Database (HMDB; http://www.hmdb.ca/) (15–17) and MassBank (http://www.massbank.jp/). Related metabolites were submitted to a correlation analysis of metabolic pathways associated with metabolic pathway analysis (MetPA) (18) and the Kyoto Encyclopedia of Genes and Genomes (KEGG; http://www.genome.jp/kegg/) (19).

Urinary metabolic fingerprints of patients with primary dysmenorrhea (T group) and healthy controls (Z group) during the luteal regression stage are displayed in Fig. 1. Several different peak intensities were clearly detected from the typical base peak intensity (BPI), which indicates two sets of data with metabolite differences in Fig. 1.

Data were processed using multivariate statistical analysis methods (Fig. 2). In the PCA scatter plot, each point represents a volunteer sample, making it possible to visually discern the differences between the samples. In addition, anomalous samples can be identified and removed (20,21) to improve the accuracy of the model. PLS-DA (Fig. 2A) was used to identify the metabolites that differentiate patients with PD from the healthy controls. The R²Y and Q² of the PLS-DA model were 0.992 and 0.806, respectively, demonstrating the accuracy of the model. Finally, a VIP value >1 and unpaired Student's t-tests were used to identify metabolites with significant differences as potential biomarkers. Following this, a score plot (S-plot) of the PLS-DA was used to identify potential discriminatory metabolites (Fig. 2B). Metabolites that are distal from the origin and close to the vertical axis of the S-plot are the differentiating metabolites.

PLS-DA analysis VIP values with >1 ion were used to identify candidate biomarkers. Significantly different endogenous compounds during the luteal regression stage were considered to be the differentiating compounds between patients with PD and healthy controls, resulting in the identification of 10 specific biomarkers. Levels of citrulline, ornithine, androstenedione, progesterone, phytosphingosine, dihydrocortisol and 17-hydroprogesterone were significantly decreased in patients with PD compared with healthy controls during the luteal regression stage (P=0.0426, P=0.0071, P=0.0040, P=0.0359, P=0.0360, P=0.0454 and P=0.0235, respectively; Table I) and levels of sphinganine, histidine and 15-keto-prostaglandin F2α were significantly increased in patients with PD compared with healthy controls during the luteal regression stage (P=0.0136, P=0.0107 and P=0.0001, respectively; Table I).

Metabolite m/z values were used to determine the probable molecular formula using the HMDB database (15–17). Of these compounds, 4 were identified using authentic standards and 6 were identified by comparing the fragments based on their molecular ion information and MS/MS data. An example is 17-hydroxyprogesterone, which exhibited an accurate biomarker mass ([M+H]⁺ at m/z 331.2241) in the mass spectrum. In positive ion mode, the MS/MS contains the fragment ions m/z 195.1 [M+H-C₉H₁₂O]⁺ and m/z 138.0 [M+H-C₁₂H₁₇O₂]⁺. The HMDB database was also used to confirm the results (15–17).

To determine the sensitivity and specificity of the 10 biomarkers identified in the luteal regression stage, SPSS 17.0 (SPSS Inc., Chicago, IL, USA) was used to analyze the ROC curves of the biomarkers. The ROC curves were regarded as a potential diagnosis threshold for the test results, and the sensitivity and specificity were calculated and evaluated. Data were plotted as 1-specificity on the × axis and sensitivity on the y axis (Fig. 3). The ROC curve is proximal to the upper left (Fig. 3B, blue line), and the largest point boundary value of the Youden index (22) is the threshold. Therefore, the sensitivity and specificity of the test are greater, and the rates of misdiagnoses and missed diagnoses are relatively low. With regard to the plotted curve and 45° oblique linear contrast, if most of the curve coincides with the independent variable, the value is a poor predictive value for the dependent variable. If the data curves away from the 45° oblique line more than the independent variable, the value is a better predictive value for the dependent variable. In the present study the biomarkers with AUC >0.7 were considered to have an important role in patients with PD. Finally, as demonstrated in (Fig. 3A), the AUC of 6 metabolites was >0.8. Thus, 6 important biomarker candidates were identified. The combination of the 10 biomarkers' AUC at the 95% confidence interval was 0.950, indicating high sensitivity and specificity during the luteal regression stage (Fig. 3B).

MetPA (18), which uses a library of metabolic pathways from the KEGG (19), was used for metabolic pathway analyses of potential biomarkers (Fig. 4). The abscissa is the impact of the pathways. Where the pathway impact value calculated from the pathway topology analysis is >0, that pathway is likely associated with PD. It was revealed that steroid hormone biosynthesis, sphingolipid metabolism, arginine and proline metabolism, histidine metabolism and arachidonic acid metabolism were perturbed at the luteal regression stage in patients with PD (Fig. 4). Finally, the disturbed metabolic pathways detected by UPLC-Q/TOF-MS analysis were analyzed (Fig. 5).