3408 - Prometheus: A Risk Classification Deep Learning Model for Predicting Liver Metastases from Pancreatic Cancer

Y. Liu¹, S. Zheng¹, P. T. Teo², M. Abazeed¹, and J. P. Hayes¹; ¹Department of Radiation Oncology, Northwestern University Feinberg School of Medicine, Chicago, IL, ²Department of Radiation Oncology, Northwestern University, Feinberg School of Medicine, Chicago, IL

Purpose/Objective(s): Some patients with pancreatic ductal adenocarcinoma (PDAC) have a very high predilection for developing early metastases. Local therapy, comprising either surgery or high-dose radiation to the pancreas, can be quite effective in preventing or delaying local progression. However, the competing risk of early distant metastases and/or death reduces the established benefits of local tumor control. A risk classification schema based on genetic biomarkers can substantially improve patient selection for local treatments. Materials/

Methods: PDAC cases from the Memorial Sloan Kettering-Metastatic Events and Tropisms (MSK-MET) dataset (n= 1779) were analyzed. PDAC cases with metastases to liver (n= 1060), biliary tract (n= 477), bowel (n= 451), lung (n= 441), distant lymph node (LN) (n= 241), and bone (n= 230) were identified. Genetic alterations including mutations and copy number variations (CNVs) were compared between and across samples with liver metastases and those from non-liver metastases. The H2O AutoML framework was used for feature evaluation and model selection. The best performing deep learning model for predicting the development of liver metastases, Prometheus, contained 8 inputs: tumor mutation burden (TMB), fractional of genome altered (FGA), mutation status of TP53, and CNVs of CDKN2A, AKT2, MYC, KRAS, and SMAD4. The MSK-MET and the independent MSK-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT, n = 178) datasets were utilized to validate the value of the Prometheus model in PDAC prognosis.

Results: Liver metastases are associated with worse prognosis (Hazard Ratio, HR=2.35, p<0.001) in multivariate analyses adjusting for metastatic categories: liver, biliary tract, bowel, lung, distant lymph node, and bone metastasis. Liver metastasis is associated with higher frequency of TP53 mutation, AKT2/MYC/KRAS CNV gains, and CDKN2A/CDKN2B/SMAD4 CNV losses. The Prometheus Score classified PDAC patients into high (Prometheus Score =70), intermediate (50=Prometheus Score<70) and low (Prometheus Score < 50) risk groups with median OS of 9.36, 15.34 and 21.39 months. The rates of liver metastasis in each group were 80.5% 61.7%, and 43.0%, respectively (p<0.001). The Prometheus model was also able to classify PDAC cases into high and low risk groups in the independent MSK-IMACT dataset with median OS of 13.41 and 26.36 months, respectively (p=0.004).

Conclusion: We successfully developed and validated a model that predicts the development of liver metastases, and overall survival in PDAC. Prometheus is based on two large cohorts and utilizes a feature reduction and optimization methodology. This model is poised for enhanced selection of patients who are more likely to benefit from local therapy to the pancreas.