Machine learning and statistical prediction of overall survival (OS) from pre-dose plasma biomarkers in a randomized phase 2 trial (1801 Part 3B) of the GSK-3 inhibitor elraglusib in metastatic pancreatic ductal adenocarcinoma (mPDAC): Application toward patient enrichment.

Abstract

4185 Background: Elraglusib is a first-in-class inhibitor of GSK-3ß, a well-credentialed target in cancer implicated in intrinsic oncologic processes and tumor immune response. Preliminary results of the 1801 Part 3B trial (NCT03678883) showed statistically significant benefits for elraglusib+GnP versus GnP for 1-year survival and mOS in mPDAC (companion abstract: Mahalingam et al.). We investigated plasma levels of cytokines/chemokines/soluble cell receptors/growth factors (CCSG) as potential biomarkers of favorable outcomes on elraglusib. Methods: Forty CCSGs were evaluated in pre-dose plasma from patients with previously untreated mPDAC enrolled in 1801 Part 3B treated with GnP (n = 78) or elraglusib+GnP (n = 155) using a Luminex immunoassay. Using Kaplan-Meier statistics and cutpoint determination, CCSGs were assessed for OS predictive ability in the elraglusib+GnP arm. Multivariate models were constructed to predict binarized survival at 12 months using machine learning (ML). Fivefold cross-validation was used in both analyses, and all methods were applied to the GnP arm to identify elraglusib+GnP specific predictors. Results: Data shown as of November 15, 2024. Multiple CCSGs significantly stratified elraglusib+GnP patients. The most extreme prognosticators were IFN-ß (average HR = 2.34), and PD-L1 (average HR = 0.52) (HR shown as high CCSG vs low CCSG). Screening of different ML approaches ranked logistic regression at the top, and hyperparameter grid search identified stochastic gradient solver with ridge regularization as the optimal method. The model had an accuracy of 88% (SD: 3.9%) and a balanced accuracy of 80% (SD: 8.5%). IFN-beta had the most substantial effect size (odds ratio (OR) = 72.28), followed by IL-18 (OR = 23.38) and PD-L1 (OR = 15.32). All other CCSGs had an OR between 0.03 and 6.71. When this model was applied to patients in the GnP arm, accuracy was 68.4%, and balanced accuracy was 43.1%. Conclusions: Many CCSG biomarkers were identified as promising predictors of survival benefit in mPDAC patients treated with elraglusib+GnP. Both univariate statistical and multivariate ML approaches show predictive significance with high interpretability. The initial ML model is specific to elraglusib treatment, not GnP alone. These results indicate that patients’ initial immune state plays a role in response to elraglusib+GnP. However, single CCSGs for enrichment currently exclude too many patients ( > 75%), and thus, panels of biomarkers are being investigated with ML to overcome this limitation. The 1801 Part 3B clinical trial recruitment has been completed, and updated biomarker models reflective of topline OS data, available by April 2025, will be presented.