Metabolomic profiling of dried blood spots reveals gender-specific discriminant models for the diagnosis of small cell lung cancer

Subjects characteristics

The study flow was shown in Figure 1. In the discovery phase, a total of 114 subjects were recruited including 40 patients with NSCLC, 37 patients with SCLC and 37 noncancer controls (Table 1). In the validation set, 78 cases were enrolled including 20 patients with NSCLC, 31 patients with SCLC and 27 noncancer controls (Supplementary Table 1). The subjects were age and sex-matched and there were no significant differences for gender, smoking status and alcohol drinking record between groups. All the lung cancer patients were at the initial diagnosis prior to any treatment. The details for clinical characteristics were listed in Table 1 and Supplementary Table 1.

The metabolomic profiling of SCLC in the DBS of the discovery set

To investigate the unique metabolic features of SCLC, we first performed comparative metabolomic profiling of DBS for the discovery cohort. After quality control analysis, we detected and identified 372 metabolites with less than 20% miss values in all the samples. The representative chromatogram from one of the DBS specimens is shown in Supplementary Figure 1. Multivariate analysis showed the metabolomic profile of SCLC in DBS was not separated from NSCLC and noncancer controls by PLS-DA when all the male and female samples were combined for analysis (Supplementary Figure 2A). In addition, ROC curve analysis revealed low diagnostic accuracy for SCLC if the males and females were analyzed together (Supplementary Figure 2B). Previous studies reported that women have a different incidence of SCLC compared to men [16, 17]. These results suggested that there might be gender differences in metabolomic profiles between male and female SCLC patients. We then analyzed the male and female samples separately.

Both 2-D and 3-D PLS-DA plots showed that male SCLC had a unique metabolic signature that was distinct from patients with NSCLC and noncancer controls (Figure 2A and 2B). Since PLS-DA is a supervised model, we then performed leave-one-out cross-validation (LOOCV) to evaluate the level of overfitting for the PLS-DA model (Supplementary Figure 3A). The Q² value of 0.76 indicated the PLS-DA model is a good predictive model for the dataset. The top 15 most discriminant metabolites revealed by variance in projection (VIP) analysis after PLS-DA was listed in Figure 2C. The VIP metabolites were further confirmed by univariate one-way ANOVA analysis. Pathway analysis revealed that sphingolipids metabolism, phospholipids, purine, urea cycle, acylcarnitines, amino acids, and endocannabinoids metabolism were significantly altered in patients with SCLC compared to NSCLC and controls (Figure 2D).

Similarly, there was a clear separation of metabolomic features in DBS among female SCLC, NSCLC and control subjects revealed by both 2-D and 3-D plots (Figure 3A and 3B). The PLS-DA classification model for the female dataset was confirmed by LOOCV (Supplementary Figure 3B). The top 7 most differential metabolites in the VIP analysis were from the biochemical pathways of phospholipids (Phosphatidylethanolamine, PE (18:2/20:4)), one-carbon metabolism (5-methyltetrahydrofolic acid), eicosanoids (9-HETE), sphingolipids (Ceramide (d18:1/20:4)), acylcarnitines (Oleoylcarnitine), branch chain amino acids (Arginine) and urea cycle (Glutamate) (Figure 3C and 3D).

The unique metabolic signatures in the DBS of male and female SCLC

We then compared the shared and unique metabolic features in male and female SCLC. As shown in the Venn diagram (Figure 4A), endocannabinoids, purine and bile acids metabolism were only significantly altered in male SCLC, while, eicosanoids, steroids, pyrimidines, and one-carbon metabolism were disturbed in female SCLC patients. Five disturbed metabolic pathways including sphingolipids, phospholipids, acylcarnitines, amino acids, and urea cycle were common to both males and females with SCLC (Figure 4A).

Even though sphingolipids and phospholipids were both altered in male and female SCLC patients, further analysis showed gender-specific differences in the subclasses of these two classes of lipids. In detail, the levels of total 2-hydroxy ceramides in DBS of male SCLC were significantly higher than male NSCLC and controls (Figure 4B). But there were no significant differences in 2-hydroxy ceramides between female subjects. The changes in the DBS of female SCLC were mainly unsaturated ceramides such as Ceramide (d18:1/20:4) (Figure 3C). Additionally, in the subclasses of phospholipids, the unsaturated PE lipids decreased dramatically only in the DBS of female SCLC (Figure 4C), while, the saturated phosphatidylinositol lipids increased significantly in male SCLC (Figure 2C).

The discriminant models of SCLC for male and female patients based on the discovery set

We next aimed to build the classification models that could be used for the identification of SCLC. We first evaluated the feasibility of the univariate model using male samples. Surprisingly, we found that a single analyte (2-arachidonylglycerol, 2-AG) had an accuracy of 85.6% (95% CI: 72.4%-95.1%) for the diagnosis of SCLC in the discovery cohort (Supplementary Figure 4A). However, it performed poorly in the validation cohort (Supplementary Figure 4B). This suggested that single metabolic biomarker might be biologically implausible for complex diseases like cancer.

We then explored the utility of the multiple variables classification. A set of 3-11 metabolites were selected and tested according to the following criteria: (1) Top 30 VIP metabolites in PLS-DA model (2) Covering as many as different biochemical pathways and k-mean clusters (3) High frequency in the least absolute shrinkage and selection operator (LASSO) model. By using this approach, the biological variation across the subjects is more readily accommodated. The performance of each classifier set of metabolites was evaluated by ROC analysis. The optimal variable combinations were obtained using the forward selection and backward elimination methods [18].

As a result, a 5-metabolite panel including 2-AG, cholic acid, PI (18:0/18:0), IMP and Cer (d18:1/22:0 OH) was selected as the best variable model for male SCLC. The biochemical pathways, VIP scores in the PLS-DA model, k-mean clusters and frequency of LASSO model of the selected metabolites were listed in Supplementary Table 2. The relative levels of PI (18:0/18:0) and Cer (d18:1/22:0 OH) were significantly upregulated in male SCLC (Figure 5A, 5B). In contrast, 2-arachidonylglycerol, IMP and cholic acid decreased significantly in the DBS of male SCLC subjects (Figure 5C–5E). The ROC analysis using the 5-metabolite panel yielded the diagnostic accuracy of 0.95 (95% CI: 0.83-1) for the diagnosis of male SCLC patients in the discovery set (Figure 5F). The selected 5-metabolite classifier was further validated by a random permutation test (500 times). A P value of 0.0084 indicated the classifier was robust (Figure 5G). The relative importance of each metabolite in the classification model of SCLC calculated by multinomial logistic regression analysis is listed in Supplementary Table 4. Among 5 selected metabolites in the model, Cholic acid has the highest odds ratio of 2.48 (95% CI: 2.19 - 2.76) for SCLC over the controls and NSCLC. In addition, multiple linear regression analysis showed a significant correlation between the serum levels of NSE and the relative levels of 5 diagnostic metabolites in the DBS of male SCLC subjects (R² = 0.63, P < 0.01, Figure 5H).

The discriminant model for female SCLC was established on the basis of another set of metabolites including PE (18:1/20:4), 5-Methyltetrahydrofolic acid, Desmosterol, 4, 5-Dihydroorotic acid, and 9-HETE. There biochemical pathways, VIP scores in the PLS-DA model, k-mean clusters and frequencies of the LASSO model were shown in Supplementary Table 3. The levels of PE (18:1/20:4), 5-methyltetrahydrofolic acid, Desmosterol, and 4, 5-Dihydroorotic acid were dramatically reduced in female SCLC compared to the controls and NSCLC (Figure 6A–6D). In contrast, 9-HETE was significantly higher in female SCLC (Figure 6E). In the discovery set, the ROC analysis using the selected analytes in DBS had the diagnostic accuracy of 0.94 (95% CI: 0.74-1) for the diagnosis of female SCLC patients (Figure 6F). Classifier robustness for the model was verified with a random permutation test with a P value of 0.0066 (Figure 6G). We then performed the multinomial logistic regression analysis to elucidate the relative importance of each metabolite for the classification of SCLC in females (Supplementary Table 5). We found that among five selected metabolites in the female model, the level of Desmosterol in DBS of SCLC has the highest an odds ratio (3.35, 95% CI: 3.06 - 3.63) over the controls and NSCLC. In addition, similar to the male classifier, a significant correlation with NSE levels in the serum was found by multiple linear regression analysis (R² = 0.56, P < 0.05, Figure 6H).

Validation of the discriminant models

To validate the accuracy of the developed models, we conducted the metabolomic profiling of another 78 DBS specimens (Supplementary Table 1). For the male subjects, the discriminant model established from the discovery set distinguished male SCLC group from those of in the validation set with the sensitivity of 86.6% (95% CI: 59.5%-98.3%) and the specificity of 84.1% (95% CI: 63.9%-95.5%) (Supplementary Table 6). Similarly, for the female subjects, the selected 5 metabolites from the discovery cohort had the sensitivity of 86.4% (95% CI: 65.1%-97.1%) and the specificity of 84.2% (95% CI: 68.8%-97.1%) for distinguishing female SCLC from the controls and NSCLC in the validation set (Supplementary Table 6).

Subjects and sample collection

This study was registered at the China Clinical Trial Registration Center (ID: ChiCTR-DDD-17010893) (http://www.chictr.org.cn/showproj.aspx?proj=18317). The study protocol was approved by the Ethics Committee of Shengjing Hospital of China Medical University on February 22, 2017. The study was conformed to the rules of the Declaration of Helsinki of 1957: Ethical Principles for Medical Research Involving Human Subjects. Written informed consent was obtained from all the subjects after full disclosure of the study details.

Inclusion criteria: volunteers in the control group met the following criteria: (1) no significant diseases (cancer, diabetes, cardiovascular disease, etc.), or recovery for more than 3 months; (2) no long-term medication history; (3) no surgery in the last 3 months; (4) age and gender-matched with SCLC patients. Patients in the SCLC and NSCLC group met the following criteria: (1) at the initial diagnosis prior to the treatment; (2) a pathological diagnosis of SCLC or NSCLC. In the discovery stage, a total of 114 subjects were enrolled and in the validation stage, another 78 subjects were recruited. The subjects’ characteristics were listed in Table 1 and Supplementary Table 1.

Blood collection and DBS preparation: the finger-prick blood was directly spotted onto Whatman 903 Specimen Collection Paper Cards (GE Healthcare, USA). Three spots were taken from each subject and the spots were dried completely and stored at room temperature with desiccant before analysis. All the specimens were fasting blood samples.

Extraction of metabolites from DBS

The extraction of metabolites from DBS was conducted according to a previous study with modifications [43]. Briefly, a 3 mm disc per subject was obtained using a GE Healthcare Uni-Core punch and transferred to a 1.5 mL tube. A blank disc from the adjacent of DBS from the same card was used as the analytical blank. One hundred ninety (195) μL of extraction buffer (isopropanol/acetonitrile/water, 3:3:2, v/v/v) and 5 μL of internal standard mix (custom-synthesized stable isotope cocktail) [44] were added to each tube. The mixture was vortexed thoroughly and incubated on ice for 10 min. Samples were centrifuged at 16,000 g for 10 min at 4 °C. The supernatant was stored at -80 °C prior to metabolomic analysis.

LC-MS/MS-based metabolomic analysis

The metabolomic analysis was performed on a Shimadzu LC-20AD (Shimadzu, Japan) coupled with a Qtrap 5500 mass spectrometer (SCIEX, USA) as described before with slight modifications [44, 45]. Briefly, the chromatographic separation was achieved on a Luna NH2 column (250 × 2 mm, 5 μm, Phenomenex, USA) with the following conditions: mobile phase A: 95% H₂O + 5% acetonitrile + 20 mM ammonium hydroxide + 20 mM ammonium acetate (pH 9.4); B: 100% acetonitrile; the flow rate: 0.35 mL/min, the column temperature: 25 °C. The injection volume was 10 μL. The gradient was: 0-3 min 95% B, 3-6 min 75% B, 6-7 min 0% B, 7-15 min 0% B, and 15-17 min, 95% B. MS/MS analysis was performed by scheduled multiple reaction monitoring (sMRM) with electrospray ionization (ESI) and positive and negative switch mode. The polarity switch time was set at 50 ms. The ion source temperature was 500 °C. A total of 420 metabolites covering the major metabolomic pathways were targeted.

The chromatographic peaks were integrated using MultiQuant 3.0 (SCIEX, USA) and confirmed by manual inspection. All the detected metabolites were normalized using the corresponding internal standards and the AUC ratios (AUC of the detected metabolites/AUC of the internal standards) prior to the statistical analysis.

Statistical analysis

The general statistical analysis was performed using GraphPad Prism 8.0 (GraphPad, USA). Data were mean and standard deviation (SD) for continuous variables and actual values with percentages for categorical variables. One-way ANOVA followed by Tukey’s test was used for multiple comparisons of continuous variables and two proportion z test was used for the categorical variables. Multiple linear regression analysis was performed to analyze the relationship between serum NSE and the selected metabolites biomarkers. Significance was set at P < 0.05.

Metabolomic analysis was performed using metaboanalyst 4.0 (https://www.metaboanalyst.ca). The k-nearest neighbor method was used to deal with missing data due to individual differences, and the metabolites with missing data > 50% were removed. The data was log 2 transformed and auto-scaled. Multivariate analysis was conducted using PLS-DA. Since PLS-DA is a supervised model, the accuracy of the PLS-DA model was verified by LOOCV and Q² values were reported which represented the estimate of the predictive ability for the models. The top 30 most influential metabolites were obtained by VIP. VIP scores > 1.5 were considered statistically significant. The metabolites were also verified by a one-way ANOVA test. The significant metabolites identified by VIP analysis was then used for biochemical pathway analysis. In addition to the PLS-DA model, both the unsupervised k-mean clustering model and the LASSO frequencies were also performed.

Sets of 1-11 metabolites were manually selected by bootstrap resampling using the VIP lists and pathways obtained PLS-DA model, k-mean clusters and LASSO frequencies as an example of one possible multianalyte diagnostic classifier. The diagnostic performance of the selected classifiers was evaluated using ROC implemented in Metaboanalyst 4.0. In order to produce a smooth ROC curve, 100 cross-validations (CVs) were performed and the results were averaged to generate the plots. Classifier robustness was estimated by the permutation test (500 times). The sensitivity, specificity, accuracy, positive predictive value (PPV) and negative predictive value (NPV) of the selected biomarkers were analyzed by 2 × 2 contingency table analysis. The relative importance of each metabolite in the predictive models was evaluated using multinomial logistic regression analysis. The multinomial logistic regression coefficients, P values, and odds ratios were reported.