Abstract 2273: A machine learning approach to identify a peripheral prognostic cytokine signature via nivolumab clearance in patients with advanced melanoma

Abstract

Abstract Background: Data suggest that peripheral factors such as circulating cytokines could function as predictive indicators of clinical response to specific therapies or as general prognostic indicators of outcome in patients with cancer. Previous studies have focused on evaluating the association between individual cytokines and clinical outcome. There is no clear consensus on how or which of these cytokines should be selected for prediction of treatment outcome. Clearance of immunotherapeutic agents such as the anti-PD-1 nivolumab (NIVO) has been shown to be a predictor of best overall response and overall survival (OS) across multiple indications. Determination of clearance requires post-treatment pharmacokinetic (PK) samples which negates its utility as a baseline prognostic factor. We investigated a novel machine learning approach to identify a baseline composite cytokine signature that correlates with the established efficacy response relationship for NIVO clearance in advanced melanoma. This approach, which is capable of integrating multivariate factors into a composite feature, may provide a more accurate reflection of underlying biologic factors that contribute to outcome following treatment. Methods: Peripheral serum PK (NIVO clearance) and pharmacodynamic (Myriad Rules Based Medicine-customized inflammatory cytokine panel) data from 2 phase 3 studies (NCT01721772; NCT01844505) of NIVO monotherapy in patients with melanoma (n = 471) were used for machine learning model development (training data set). A third phase 3 study of NIVO in advanced melanoma (NCT01721746) was included in model application (test data set; n = 158). Random forest, a tree-based ensemble learning method, was used for feature selection and classification (low vs high) of NIVO clearance. Kaplan-Meier analyses and log-rank p values were used to assess statistical difference in long-term OS between predicted high- vs low-clearance groups. Results: A panel of ~20 baseline inflammatory cytokines related to immune cell modulation were selected by the machine learning model to predict clearance of NIVO monotherapy in patients with advanced melanoma (area under the curve of receiver operating characteristic: 0.79). The predicted clearance value from the cytokine signature was significantly associated with OS across all 3 studies, including patients treated with NIVO or dacarbazine (control) (p<0.01). Conclusions: We developed a machine learning approach to identify a serum prognostic cytokine signature based on the relationship between NIVO clearance and response. The strong association of the defined cytokine signature with OS from NIVO and dacarbazine suggests a prognostic role for drug clearance. This cytokine signature has the potential to be a prognostic biomarker in melanoma and may be used for balancing and randomization of patients in clinical studies. Citation Format: Rui Wang, Xiao Shao, Junying Zheng, Abdel Saci, Max Qian, Irene Pak, Amit Roy, Akintunde Bello, Jasmine Rizzo, Fareeda Hosein, Rebecca A. Moss, Megan Wind-Rotolo, Yan Feng. A machine learning approach to identify a peripheral prognostic cytokine signature via nivolumab clearance in patients with advanced melanoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 2273.