Identification of Serum Markers of Esophageal Adenocarcinoma by Global and Targeted Metabolic Profiling

Study design

This is a retrospective study in which two sets of case-control pairs, each consisting of 30 cases and 30 controls, were included in discovery phase for global metabolomic profiling and screening. Validation phase comprised 321 EAC cases and 331 controls. Cases and controls in both phases were frequency-matched on age and gender and were all Caucasians. Cases with histologically confirmed EAC and no prior chemotherapy or radiotherapy were recruited between January 2004 and December 2014 from The University of Texas MD Anderson Cancer Center through a daily review of computerized appointment schedules. There was no age, gender, or clinical stage restrictions. Controls were selected from general population, who had no prior history of cancer except non melanoma skin cancer and matched to cases by age, gender, ethnicity, and residence. The controls were identified and recruited using the random digit dialing method. The same questionnaire data were collected from the cases and controls, which included demographic and epidemiologic risk factor data for EAC, including smoking, obesity, and medical history of GERD and Barrett’s esophagus. BMI (kg/m2) was categorized according to the standard classification of the World Health Organization (non-obese or normal/overweight<30; obese >=30). An individual who never smoked or smoked no more than 100 cigarettes in his or her lifetime was considered a never smoker. An individual who smoked at least 100 cigarettes in his or her life time was defined as an ever smoker. Individuals with stage I, IIa and IIb were classified as early-stages whereas individuals with stage III, IVa and IVb were categorized as late-stages. All study participants signed a written informed consent, and the study was approved by the Institutional Review Boards of The University of Texas MD Anderson Cancer Center (Houston, TX).

Global metabolic profiling and individual targeted metabolomic validation

Global metabolomic profiling analyses were conducted with 200 ul of serum samples at Metabolon Inc. (Durham. NC) as previously described⁹. Individual targeted metabolomic validation was carried out for the three individual selected metabolites in validation dataset using 100 ul of serum for each metabolite by using validated LC-MS/MS methods at Texas Southern University. The LC-MS/MS assays were validated according to FDA’s requirements described in its ‘Guidance for Industry: Bioanalytical Method Validation’¹⁰. Our LC-MS/MS assay met FDA’s validation requirements. In addition, other criteria such as selectivity, recovery, matrix effect, stability, and reproducibility of the assay were also well demonstrated¹¹.

Pathway analyses and individual metabolite selection

For pathway identification of the altered metabolites, the Kyoto Encyclopedia of Genes and Genomes (KEGG) database¹² was used. EAC-associated metabolites were mapped to metabolic maps using their compound IDs, followed by identification of the anabolic and catabolic enzymes in the mapped pathways. To delineate the overall biological processes altered in EAC, we examined whether EAC-associated metabolites and their corresponding mapped genes by using the Enrich analysis program ¹³. We then used a pathway-based selection approach for identifying top metabolic markers. Three of the top differentially expressed metabolites in both discovery datasets were selected and analyzed by targeted metabolomic profiling in the validation phase based on their biological relevance, differential expression levels, and validated assay availability.

Statistical Methods

Statistical analyses were performed using Stata (version 10.1; StataCorp). Differences in distribution of the host characteristics between cases and controls were evaluated by Pearson χ² test for categorical variables whereas Student t-test was used to test differences for continuous variables.

For global metabolomic profiling analysis of the case-control series in discovery phase, data were log2 transformed. Only metabolites with detectable expression in at least 80% of samples were kept for analysis. A total number of 330 metabolites from Discovery I dataset and 341 from Discovery II dataset were included for analyses. Wilcoxon-rank sum test was conducted to test the difference between cases and control as overall mean values and in stratified analyses by sex, age (<63 vs >=63), smoking status (Never vs. Ever), BMI(<30 vs >=30), and to test the difference between late and early stage cases. Association between LP, BHBA and DM with EAC risk was estimated by logistic regression analyses in validation dataset to determine the odds ratio (OR) and 95% confidence interval (95% CI) adjusting for age, gender, smoking status, BMI and batch effect.

Receiver operating characteristic (ROC) curves were used to determine the value with the highest sensitivity and specificity of individual metabolites to be used as cutoff point for dichotomizing patients. Individual metabolites were also analyzed as categorical variables by setting cutoff points at quartile values in control group. In addition, stratified analyses according to age, gender, smoking status and BMI were performed. We then added an interaction term to the logistic regression models using median as cutoff point for a better data distribution to test the interaction between the metabolites and smoking status and BMI in modulating EAC risk. The combined metabolite risk score (MRS) for each patient was derived by linear combination of the product of reference-normalized level of each of the three metabolites by its logistic regression corresponding coefficient. All cases were dichotomized by the using the ROC method into high or low risk groups, respectively. We also added an interaction term to the logistic regression models to test the interaction between the MRS and BMI and smoking in modulating EAC risk. All statistical tests were two sided, with statistical significance set at P<0.005.

Patient characteristics

We included 30 cases and 30 controls in each of the two discovery phases for screening and identification of promising metabolites. The validation phase consisted of 321 cases and 331 controls who were frequency-matched on age and gender (Table 1). Mean age was 62.99±10.69 years in cases and 62.88±10.33 in controls (P=0.890). Over 90% of cases were males, consistent with the EAC male dominance. As expected, cases had significantly higher proportion of ever smokers than controls (73.89% vs. 51.96%, P<0.001). Among ever smokers, self-reported median pack-years were significantly higher in cases than in controls (6.14% vs. 36.91%, P<0.001). There was no significant difference in neither BMI nor diabetes status between cases and controls. Among cases in the validation set, 146 subjects were early-stage patients while 175 were late-stage.

Differentially expressed metabolites in the discovery phase

A total of 437 metabolites were detected in the discovery phase. After quality control analyses, 331 metabolites (Discovery I) and 341 metabolites (Discovery II) were included for analyses (Supplementary Figure 1). Wilcoxon rank-sum test identified 131 metabolites in Discovery I and 123 metabolites in Discovery II as showing significant differential levels between cases and controls (p<0.005). Taken together, 64 metabolites had consistent differential levels between cases and controls in both discovery batches (P<0.005). Out of these 64 metabolites, 22 had significantly higher levels in cases compared to controls whereas 42 exhibited lower levels (Supplementary Table 1).

Pathway enrichment analyses of EAC-associated metabolites

To explore the relationships between these metabolic biomarkers, their corresponding metabolic pathways were analyzed. These metabolites were mapped to 5 super-pathways (amino acids, carbohydrates, lipids, peptides and xenobiotics) and 27 sub-pathways revealing an increase in amino acid metabolism and changes in carbohydrates, lipids and gamma-glutamyl aminoacids as potential pathways involved in EAC (Supplementary Table 2). A similar enrichment network was identified by the bioinformatics tool Enrich¹³ pointing to alterations in protein biosynthesis, fatty acids, energy and bile acid metabolism as key metabolic EAC-pathways, with Arginine and Proline metabolism identified as the main significantly overrepresented pathway (P<0.001; FDR<0.01) (Supplementary Figure 2).

Individual targeted validation of selected metabolites

Metabolites that exhibited a significant trend in altered levels from normal individuals to early- to late-stage EAC cases are potential early detection biomarkers of EAC. Out of the 64 significant metabolites, we selected three metabolites, an amino acid LP, a ketone body BHBA, and a carbohydrate DM based on their differential expression levels, biological significance, and previous literature reports. We developed validated standard LC/MS-MS assays and measured their levels in 321 EAC cases and 331 matched controls for validation. Consistent with discovery phase, cases had lower levels of LP than controls (mean±SD: 22.78±6.79 vs. 28.24±8.64 ug/ml; P<0.001) and higher levels of BHBA (18.06±17.84 vs. 7.73±9.92 ug/ml; P<0.001) and DM (9.87±4.28 vs. 6.28±3.61 ug/ml; P<0.001) (Table 2). Results of stratified analyses by gender, age, smoking status and BMI were consistent with overall results (Supplementary Table 3).

Association of serum metabolite levels with EAC stage

Since clinical stage is an established EAC prognostic factor EAC, we compared the levels of these 3 metabolites between controls and patients with early-stage and late- stage in validation set. Controls had higher levels of LP and lower levels of BHBA and DM when compared to early-stage or late-stage cases separately (P<0.001 for all comparisons).

When comparing early- to late-stage patients, no significantly different levels were found for LP. A borderline significance was found for BHBA with higher levels in late-stage cases when compared to early-stage (mean±SD, 19.89 ± 18.97 vs 15.89±16.19 ug/ml, P=0.071) and significantly higher levels when compared to controls (mean± SD, 19.89±18.97 vs 7.73±9.92 ug/ml, P<0.001). Interestingly, levels of DM were significantly higher in late-stage patients when compared to early-stage patients (mean±SD, 10.61±4.79 vs. 8.97 ± 3.36 ug/ml; P=0.005) and controls (10.61 ± 4.79 vs. 6.28 ± 3.61 ug/ml; P<0.001) (Table 3).

Associations between LP, BHBA and DM and EAC risk

We next analyzed the association between the 3 validated metabolites with EAC risk by unconditional multivariate logistic regression analyses in validation set. Using the ROC method to determine the cutoff point of each metabolite, higher levels of LP were associated with a significant decreased EAC risk (OR=0.26; 95% confidence interval (CI), [0.18–0.38]; P<0.001). However, a significantly increased EAC risk was associated with higher levels of either BHBA (OR =4.05, 95% CI, 2.84–5.78) or DM (OR =7.04, 95% CI, 4.79–10.34). Quartile analyses showed a dose-response relationship. Using the lowest quartile as reference, the ORs for the 2^nd, 3^nd and 4^th quartiles were 0.41 (95% CI, 0.26–0.64), 0.28 (95% CI, 0.18–0.45) and 0.14 (95%CI, 0.08–0.25), respectively (P for trend<0.001) for LP. In contrast, using the lowest quartile as reference, the ORs for the 2^nd, 3^th and 4^th quartiles were 2.97 (95% CI, 1.55–5.70), 4.06 (95% CI, 2.15–7.70) and 9.45 (95% CI, 5.15–17.34), respectively, (P for trend< 0.001) for BHBA, and 2.54 ( 95% CI, 1.21–5.31), 4.26 (95% CI, 2.11–8.60) and 16.97 (95% CI, 8.63–33.37), respectively, (P for trend<0.001) for DM (Figure 1). Stratified analyses by sex, age, smoking status and BMI showed consistent results with overall analyses (Supplementary Table 4).

We also evaluated the joint effects of these metabolites with smoking status and BMI, two known EAC risk factors. A joint effect between low LP levels and smoking was seen. On the other hand, a joint effect was observed between either both high BHBA and DM levels and smoking. In terms of BMI, a joint effect between low LP levels and BMI (>=30) was also seen. A similar joint effect was seen for both high BHBA and DM levels and BMI (Table 4).

Development of a metabolite risk score

The combined effects of the three validated metabolites were investigated by calculating a metabolite risk score (MRS). When using ROC curve to determine the cutoff point, patients with high MRS exhibited a significantly increased EAC risk compared with those with low MRS (OR=7.76, 95%CI, 4.80–12.53; p<0.001). A joint effect between high MRS and smoking status was seen, with OR of 11.06 (95% CI, 5.05–24.19) and 4.14 (95% CI, 1.83–9.37) for individuals with one risk factor and OR of 23.40 (95% CI, 10.95–50.00) for individuals with both risk factors (smoking and high MRS) with a significant interaction (P interaction <0.001) (Figure 2).