A Machine‐Learning Approach to Identify a Prognostic Cytokine Signature That Is Associated With Nivolumab Clearance in Patients With Advanced Melanoma

Abstract

Lower clearance of immune checkpoint inhibitors is a predictor of improved overall survival (OS) in patients with advanced cancer. We investigated a novel approach using machine learning to identify a baseline composite cytokine signature via clearance, which, in turn, could be associated with OS in advanced melanoma. Peripheral nivolumab clearance and cytokine data from patients treated with nivolumab in two phase III studies (n = 468 (pooled)) and another phase III study (n = 158) were used for machine‐learning model development and validation, respectively. Random forest (Boruta) algorithm was used for feature selection and classification of nivolumab clearance. The 16 top‐ranking baseline inflammatory cytokines reflecting immune‐cell modulation were selected as a composite signature to predict nivolumab clearance (area under the curve (AUC) = 0.75; accuracy = 0.7). Predicted clearance (high vs. low) via the cytokine signature was significantly associated with OS across all three studies (P < 0.01), regardless of treatment (nivolumab vs. chemotherapy).

Confounding factors in exposure–response analyses and mitigation strategies for monoclonal antibodies in oncology

“…Model-predicted clearance via cytokine signature was significantly associated with OS across all of the studies, which further supports the hypothesis for the relationship between clearance and disease status. 59,60 Looking prospectively, these collective observations also suggest that the presence of timedependent PK, and the significance of albumin or tumour burden as a covariate on CL would indicate the risk of a confounded E-R analysis.…”

Application of Machine Learning for Tumor Growth Inhibition – Overall Survival Modeling Platform

“…[27][28][29][30] The reported use of ML for survival outcome in oncology has been few and typically limited to using high-dimensional imaging or gene expression data as predictors, [31][32][33] and recently ML was applied to identify the association between baseline biomarker signature and nivolumab clearance, which is linked to survival outcome. 34 An evaluation by the FDA of simulated data showed that ML-based methods outperformed Cox model in survival prediction performance and in identifying the preset influential variables, and the authors of that analysis concluded that ML-based methods provide a powerful tool for time-to-event analysis, due to their capacity for high-dimensional data and better performance when the predictor variables assume nonlinear relationships in the hazard function. 18 The current exploratory analysis was based on the atezolizumab study OAK, using historical patient-level data typically available from a phase III clinical trial, applying four well-established ML methods to investigate feature selection and the predictive performance of each method and to compare strengths and weaknesses among the four ML methods and the previously developed TGI-OS model applied to the same data.…”

“…The reported use of ML for survival outcome in oncology has been few and typically limited to using high‐dimensional imaging or gene expression data as predictors, 31–33 and recently ML was applied to identify the association between baseline biomarker signature and nivolumab clearance, which is linked to survival outcome 34 . An evaluation by the FDA of simulated data showed that ML‐based methods outperformed Cox model in survival prediction performance and in identifying the preset influential variables, and the authors of that analysis concluded that ML‐based methods provide a powerful tool for time‐to‐event analysis, due to their capacity for high‐dimensional data and better performance when the predictor variables assume nonlinear relationships in the hazard function 18 …”

Application of Machine Learning in Translational Medicine: Current Status and Future Opportunities

“…For example, the complex determinants of PK and efficacy of cancer immunotherapy have been described extensively for checkpoint inhibitors where increased baseline clearance of protein therapeutics, likely explained by patient-specific disease status factors (e.g., cancer-related cachexia), has been associated with poor clinical outcomes ( 40 ). A recent ML-based analysis using RF was able to identify a cytokine signature that was predictive of high baseline clearance of nivolumab and offered generalizable translatability as a predictor of overall survival, representing an example of reverse translation that was able to discover molecular determinants of important associations between baseline (patho)physiology and disease trajectory that have confounded exposure-response understanding for this class of biotherapeutics ( 41 ). Importantly, in addition to enabling such associations between patients’ response, or late clinical endpoint, and patients’ baseline factors (including demographics, clinical, genetic, laboratory, and imaging data) or early biomarkers, such algorithms offer the opportunity to integrate and assess the predictive power of large sets of longitudinal factors on the considered outcome.…”