Clinically, to make timely treatment decisions, it would be beneficial to have a noninvasive test that distinguishes a benign from a malignant pancreatic and periampullary mass or stricture. A number of investigators have demonstrated that the serum metabolomic profile can discriminate pancreatic cancer from benign pancreatic lesions [11
]. However, in the clinic, it is not always possible to distinguish pancreatic cancer from other periampullary adenocarcinomas. Therefore, our goal was to identify the metabolomic features that separated all pancreatic and periampullary adenocarcinomas from benign masses and strictures using a two-class model. This was considered feasible because cancers have some common features to their disordered metabolism, such as the Warburg effect.
We determined that focused metabolomic profiles containing as few as 14–18 metabolites do discriminate between serum samples from patients with malignant versus benign pancreatic/periampullary lesions. In the training set, these focused metabolomic profiles produced OPLS-DA models with R2
values of 0.30–0.48, indicating that 30%–48% of the observed variance in metabolite levels was attributable to the diagnostic classification. These values are in the range expected for clinical specimens [30
], are in keeping with the clustering of samples by diagnostic category seen in the first and second components of unsupervised PCA, and were sufficient to statistically discriminate between diagnostic classes as indicated by the CV-ANOVA p-values. In separate validation sets, the AUROC values of metabolomics models had a level of performance similar to that of the serum tumor marker CA 19-9, suggesting that the metabolomic profile may have some value [1
]. However, the test performance is insufficient to have a direct impact on clinical decision-making in its present form.
Our metabolomic models did not perform as well as in other studies, including our own. This is because our comparator groups consisted of disease processes that were more heterogenous than in other series, which typically consisted of pancreatic adenocarcinoma versus pancreatitis and/or or normal controls [11
]. This was by design. Clinically, pancreatic and non-pancreatic adenocarcinomas are frequently indistinguishable. (It is only after resection that they can be definitively classified.) Therefore, if one were to apply a metabolomic profile for pancreatic cancer (as described by previous investigators) to a clinical population, it would underperform because of the inherent heterogeneity of the metabolomic features of similar lesions. We wanted to see if there was a simple “adenocarcinoma profile” that might be more applicable. The inferior performance of our model was not completely surprising, given that pancreatic cancer is associated with diabetes and non-pancreatic periampullary adenocarcinomas are not. Our experimental design provides a more realistic estimate of the performance of a test based on a single two-class metabolomic model. To enhance the performance of a metabolomics-based blood test for identification of individuals with malignancy, it will be important to define the metabolomic alterations associated with pancreatic and non-pancreatic adenocarcinomas separately.
The results of previous studies may also be overly optimistic because external validation sets were not always reported. In most studies, excellent AUROCS have been reported when metabolomic models were internally validated [11
], exaggerating the performance of the metabolomic models. When external cohorts are used for validation, AUROCs on the validation sets are typically lower than in internal validation cohorts. For example, Kobayashi et al. compared metabolomic profiles of sera from pancreatic cancer and chronic pancreatitis analyzed by GC-MS [14
]. They described an AUROC of 0.93 in a training set, and the AUROC in the validation set was only 0.76, which was no better than CA19-9 (AUROC 0.82) and CEA (AUROC 0.80). This illustrates the critical need for validation sets wherever possible in this field.
Unexpectedly, combining two complimentary analytical modalities in an attempt to gain a more comprehensive interrogation of the metabolome did not markedly improve test performance. We had hypothesized that creating a combined 1H-NMR spectroscopy/GC-MS dataset could harness the relative strengths of each platform, providing stronger predictions. However, the combined dataset models performed only slightly better than the GC-MS models and not as well as the 1H-NMR models. The combined metabolomic model derived from 1H-NMR spectroscopy and GC-MS was limited by the stability of the latter. 1H-NMR spectroscopy models generally performed better than the GC-MS models, with a smaller standard error for the AUROC values. In addition, the average standard error for the metabolite coefficients was 62% higher for GC-MS compared to 1H-NMR spectroscopy, indicating more variability in the regression modeling of metabolites in the GC-MS model. Finally, the consistency of metabolites identified across allocation sets was higher for 1H-NMR spectroscopy than for GC-MS. For 1H-NMR spectroscopy, 58% of metabolites important to the final focused list were identified in two or more of the allocation sets; only 36% of metabolites in the final list were identified in two or more of the allocation sets in GC-MS (p = 0.04).
It is possible that alternative approaches will be required to merge data from two analytical platforms. In the present study, a simple averaging technique was used to combine metabolites detected by both platforms. Other approaches to combining data are being developed and used, but no standard approach has yet been established [34
]. Further work in this field may result in a method that effectively capitalizes on the relative strengths of multiplatform detection, to produce an even stronger diagnostic model.
Compared to our previous study using 1H-NMR spectroscopy to distinguish pancreatic cancer from benign tumors, some of the same changes in metabolites were observed in the present study. Specifically, in both studies, malignancy was associated with elevated glutamate, phenylalanine, and mannose; creatine, glutamine, threonine, and lysine were higher in the benign condition. These metabolites did not vary significantly by sex or any other factor other than disease state. The fact that these changes were seen in independent studies, despite the heterogeneity of the cancers included in the present study is encouraging.
The metabolomic profiles identified offer insights into the metabolomic pathways altered in patients with a pancreatic and periampullary malignancy. The results clearly indicate a tipping of the balance of amino acid metabolism towards higher glutamate levels in malignant samples and higher glutamine and alanine levels in benign disease. These observations are consistent with earlier findings published by our group, which found elevated levels of glutamate in the serum of pancreatic cancer patients when compared to that of patients with benign pancreatic or biliary disease [11
]. It is also consistent with findings in many cancer model systems that show a switch to glutamine consumption, and increased glutamate and succinate production, in patients with rapidly proliferating cancer cells, as part of the “Warburg effect” [36
Arginine and ornithine are part of the urea cycle and feed the production of putrescine, the rate-limiting step in protein synthesis. The conversion of arginine to ornithine, by the enzyme arginase, has been suggested as a major regulator of cell growth [36
]. It is therefore interesting that arginine and ornithine levels were lower in the serum of patients with pancreatic cancer compared to benign pancreatic lesions. The level of urea, a side product of arginine-to-ornithine conversion, was slightly higher in patients with pancreatic cancer. Together, these findings are consistent with altered arginase activity in patients with pancreatic and periampullary adenocarcinomas. In pancreatic cancer samples tested by the International Cancer Genome Consortium [38
], only 1% contained ARG1 mutations, and no ARG2 mutations were seen. Therefore, the mechanism for this observation must be further elucidated.
The correlation of serum galactose with pancreatic cancer is seen in the GC-MS dataset only, as NMR did not detect galactose in these conditions. A similar relationship was recently observed with colorectal cancer patients [18
], but further investigation is needed before any putative mechanism can be proposed.