Developing a Predictive Model for Metastatic Potential in Pancreatic Neuroendocrine Tumor

Abstract

Context: Pancreatic neuroendocrine tumors (PNETs) exhibit a wide range of behavior from localized disease to aggressive metastasis. A comprehensive transcriptomic profile capable of differentiating between these phenotypes remains elusive.

Objective: Use machine learning to develop predictive models of PNET metastatic potential dependent upon transcriptomic signature.

Methods: RNA-sequencing data were analyzed from 95 surgically resected primary PNETs in an international cohort. Two cohorts were generated with equally balanced metastatic PNET composition. Machine learning was used to create predictive models distinguishing between localized and metastatic tumors. Models were validated on an independent cohort of 29 formalin-fixed, paraffin-embedded samples using NanoString nCounter®, a clinically available mRNA quantification platform.

Results: Gene expression analysis identified concordant differentially expressed genes between the 2 cohorts. Gene set enrichment analysis identified additional genes that contributed to enriched biologic pathways in metastatic PNETs. Expression values for these genes were combined with an additional 7 genes known to contribute to PNET oncogenesis and prognosis, including ARX and PDX1. Eight specific genes (AURKA, CDCA8, CPB2, MYT1L, NDC80, PAPPA2, SFMBT1, ZPLD1) were identified as sufficient to classify the metastatic status with high sensitivity (87.5-93.8%) and specificity (78.1-96.9%). These models remained predictive of the metastatic phenotype using NanoString nCounter® on the independent validation cohort, achieving a median area under the receiving operating characteristic curve of 0.886.

Conclusion: We identified and validated an 8-gene panel predictive of the metastatic phenotype in PNETs, which can be detected using the clinically available NanoString nCounter® system. This panel should be studied prospectively to determine its utility in guiding operative vs nonoperative management.

Keywords: PNET; machine learning; neuroendocrine tumor; pancreatic neuroendocrine tumor.

Figure 1.

Unsupervised hierarchical clustering identifies differential expression between localized and metastatic PNETs. (A, B) PCA of RNA-Seq in PNET tumors of dataset-1 (n = 46 samples) and dataset-2 (n = 49 samples). The arrows correspond to samples that were thought to be localized at the time of diagnosis but ultimately developed a distant recurrence postoperatively. (C, D) Volcano plot shows log₂ fold change plotted against significance (–log₁₀FDR scale, Wald test). Each dot represents an individual gene differentially expressed (FDR ≤ 0.1 after Benjamini–Hochberg correction). Positive and negative log₂ changes indicate enrichment and downregulation in metastatic tumors, respectively. (E, F) Visual representation of genes overexpressed in localized tumors (E) or genes overexpressed in metastatic tumors (F) in dataset-1 and dataset-2. Of the 29 726 genes identified in dataset-1, 219 were overexpressed in localized tumors whereas 184 were overexpressed in metastatic tumors. Of the 25 837 genes identified in dataset-2, 785 were overexpressed in localized tumors and 873 were overexpressed in metastatic tumors. Twenty genes had concordant differential expression patterns (12 overexpressed in localized tumors and 8 overexpressed in metastatic tumors) between the 2 datasets. Abbreviations: FDR, false discovery rate; PCA, principal component analysis; PNET, pancreatic neuroendocrine tumor; RNA-Seq, RNA sequencing.

Figure 2.

Interrogation of differentially expressed gene sets in localized and metastatic PNETs and extraction of highly recurrent contributory genes. (A) GSEA analysis demonstrating the highest-ranked pathways significantly altered (FDR <0.05 after Benjamini–Hochberg correction) in localized and metastatic PNETs. Gene sets with concordant upregulation or downregulation in localized and metastatic tumors across both dataset-1 and dataset-2 were included. (B) Visual representation of the gene set score, defined as the sum of the normalized effect size across all the gene sets containing a specific gene. The genes that generated the top gene-set scores are represented along the y-axis. Abbreviations: FDR, false discovery rate; GSEA, gene set enrichment analysis; PNET, pancreatic neuroendocrine tumor.

Figure 3.

Feature selection and optimization. (A) Boruta feature selection was employed to identify the most informative genes for classification within the training dataset prior to model generation. Genes that contained limited unique expression values across samples or were highly correlated (R > 0.9) with others were excluded a priori. The Boruta algorithm was performed over 1000 permutations (threshold P-value <.01), and 8 genes were thereafter identified as relevant for classification or localized or metastatic phenotype, with all downstream modeling constrained to these genes. (B) Violin plots of the 8 genes identified by Boruta feature selection. Each dot represents the expression value (log₂TMP + 1) from localized (18) or metastatic (red) PNET samples. Five of the genes (CPB2, MYT1L, PAPPA2, SFMBT1, ZPLD1) were identified from the differential gene expression analysis whereas 3 (AURKA, CDCA8, NDC80) were identified by extraction from the GSEA. Abbreviations: GSEA, gene set enrichment analysis; PNET, pancreatic neuroendocrine tumor.

Figure 4.

Machine learning and construction of predictive models on RNA-Seq data. (A, B) Predictive performance of the 4 statistical models in the testing dataset. (C) Heatmap of the scaled variable importance of each of the 8 genes to the 4 statistical models (D) Prediction correlation matrix of the 4 statistical models employed estimated by pairwise Spearman correlation. Correlation ranged from 74% to 83%. (E) Probability distributions of the classifier indices of phenotypes for localized and metastatic PNETs RNA-Seq data. (F) Theoretical PPV and NPV of each test using Bayesian analysis dependent upon the prevalence of metastatic disease within the test population and each model's sensitivity and specificity. Abbreviations: NPV, negative predictive value; PNET, pancreatic neuroendocrine tumor; PPV, positive predictive value; RNA-Seq, RNA sequencing.

Figure 5.

Validation of predictive models on NanoString nCounter® platform with risk stratification of individual samples. (A) Predictive performance of the 4 statistical models in the NanoString nCounter® dataset. (B) The probability distributions of metastatic phenotype extracted from the predictive models on 29 independent FFPE-derived samples. Samples are split into those that were localized at the time of diagnosis (top) and metastatic at the time of diagnosis (bottom). Classifier index ≤ 25% was considered “low risk,” 25% to 50% “intermediate risk,” and >50% “high risk.” Two samples (2 and 3) were localized at the time of diagnosis but ultimately developed a distant postoperative recurrence. Abbreviations: FFPE, formalin-fixed, paraffin-embedded.