Abstract 5270: Developing a predictive model for pancreatic neuroendocrine tumor metastatic potential: A multi-institutional analysis

Abstract Pancreatic neuroendocrine tumors (PNETs) are rare neoplasms that arise from cells in the islets of Langerhans, with surgical resection presently recommended for tumors > 2cm. While many PNETs have the propensity to be indolent, some small tumors display aggressive features with early metastatic potential. We used machine learning to develop a predictive model of metastatic potential dependent upon the transcriptomic signature of primary PNET tissue. To build this model, RNA sequencing data was obtained from the primary tissue of 96 surgically-resected PNETs from various institutions. Two cohorts were generated with equally balanced metastatic PNET composition (15 (32.6%) vs. 13 (26.5%), p=0.52). A differential gene expression analysis identified 20 concordantly differentially expressed genes associated with metastatic status between the two cohorts. Unsupervised surrogate variable analysis estimated and adjusted for significant sources of variation not related to metastatic potential and mitigated unwanted noise and batch effects. A gene set enrichment analysis identified an additional 29 genes that most frequently contributed to the enriched biologic pathways extrapolated from the sequencing data. Log transformed, batch corrected TPM values for these 49 genes were combined with an additional 10 clinically-relevant genes, including ARX and PDX1, that are known to contribute to PNET signatures or oncogenesis. The datasets were subsequently randomized in a 1:1 ratio and informative features with respect to metastatic status were identified utilizing a Boruta algorithm, with a priori exclusion of highly-correlative genes and those that displayed near zero variance. Nine genes, including AURKA, ARX, CDCA8, CPB2, MYT1L, NDC80, PAPPA2, SFMBT1 and ZPLD1, were identified as sufficient to classify the localized or metastatic outcome. Distributed random forests (DRF), generalized linear models (GLM), gradient boosting machines (GBM) and extreme gradient boosting (XGBoost) models were trained utilizing these 9 genes. Training ROC ranged from 0.92 for DRF to 1 for XGboost. When applied to 47 independent validation samples, the testing sensitivity ranged from 75% for DRF to 94% for GBM; specificity ranged from 84% for DRF to 94% to XGboost and GLM; positive predictive value ranged from 72% for DRF to 86% for GLM; negative predictive value ranged from 88% for GLM to 97% to GBM. The degree of predictive agreement between models ranged from 64% to 91%. Taken together, we have developed a highly sensitive predictive model of the metastatic PNET phenotype that is based on expression of nine genes. Its application as a guide for management should be studied prospectively in patients with newly diagnosed PNETs. Citation Format: Jacques A. Greenberg, Nikolay A. Ivanov, Yajas Shah, Scott Kulm, Jelani Williams, Catherine G. Tran, Theresa Scognamiglio, Yeon Joo Lee, Caitlin E. Egan, Irene M. Min, Rasa Zarnegar, James Howe, Xavier Keutgen, Thomas J. Fahey, Olivier Elemento, Brendan M. Finnerty. Developing a predictive model for pancreatic neuroendocrine tumor metastatic potential: A multi-institutional analysis [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5270.