Blood-Based Diagnosis and Risk Stratification of Patients with Pancreatic Intraductal Papillary Mucinous Neoplasm (IPMN)

Abstract

Purpose: Intraductal papillary mucinous neoplasm (IPMN) is a precursor of pancreatic ductal adenocarcinoma. Low-grade dysplasia has a relatively good prognosis, whereas high-grade dysplasia and IPMN invasive carcinoma require surgical intervention. However, diagnostic distinction is difficult. We aimed to identify biomarkers in peripheral blood for accurate discrimination.

Experimental design: Sera were obtained from 302 patients with IPMNs and 88 healthy donors. For protein biomarkers, serum samples were analyzed on microarrays made of 2,977 antibodies. A support vector machine (SVM) algorithm was applied to define classifiers, which were validated on a separate sample set. For microRNA biomarkers, a PCR-based screen was performed for discovery. Biomarker candidates confirmed by quantitative PCR were used to train SVM classifiers, followed by validation in a different sample set. Finally, a combined SVM classifier was established entirely independent of the earlier analyses, again using different samples for training and validation.

Results: Panels of 26 proteins or seven microRNAs could distinguish high- and low-risk IPMN with an AUC value of 95% and 94%, respectively. Upon combination, a panel of five proteins and three miRNAs yielded an AUC of 97%. These values were much better than those obtained in the same patient cohort by using the guideline criteria for discrimination. In addition, accurate discrimination was achieved between other patient subgroups.

Conclusions: Protein and microRNA biomarkers in blood allow precise diagnosis and risk stratification of IPMN cases, which should improve patient management and thus the prognosis of IPMN patients. See related commentary by Löhr and Pantel, p. 1387.

Figure 1.

Performance of serum protein panels to distinguish IPMN patients from healthy individuals or discriminate between different malignancy risks. A, Workflow of the process for the identification of serum protein biomarker classifiers. B, Performance results are presented as ROC curves and corresponding AUC values as determined in the discovery and validation cohorts, respectively. C, For some typical proteins, the difference is shown as a boxplots, indicating median, first and third quartile as well as maximum and minimum scores.

Figure 2.

Performance of serum miRNA panels to distinguish IPMN patients from healthy individuals or discriminate between different malignancy risks. A, Workflow of the process for the identification of serum miRNA biomarker classifiers. B, Performance results are presented as ROC curves and corresponding AUC values as determined in the training and validation cohorts, respectively. C, For some typical miRNA molecules, the variations in their abundance level across IPMN grades are shown as boxplots, indicating median, first and third quartile as well as maximum and minimum scores. A, adenoma; B, borderline; C, carcinoma in situ; I, invasive carcinoma.

Figure 3.

Diagnostic performance of a combined panel of protein and miRNA biomarkers to discriminate high-risk from low-risk IPMN. The results are presented as ROC curves and corresponding AUC values as determined in the training and validation cohorts, respectively.

Figure 4.

Diagnostic performance of clinical parameters to discriminate high-risk from low-risk IPMN. The 8 patient characteristics of preoperative CA19–9 levels, diagnosis of pancreatitis, cyst diameter, thickened/enhancing cyst wall, detection of a solid component, main pancreatic duct dilatation, lymphadenopathy, and obstructive jaundice were used for the training of an SVM classifier. The training and validation cohorts consisted of 145 and 65 samples, respectively. A, The discriminative performance of the classifier is shown as ROC curves and corresponding AUC values. B, As a means to check for overfitting, also the reverse analysis was performed. The 65 samples were used for independent training, whereas validation was done on the 145 samples. Again, the resulting ROC curves and corresponding AUC values are shown.