Abstract 2782: Machine learning algorithm identifies key serum cytokines associated with evidence of clinical activity in patients treated with personalized immunotherapeutic platform (IGV-00)

Abstract

Abstract Background: IGV-001 is a novel immunotherapy that combines irradiated, patient-derived glioblastoma tumor cells and an antisense oligonucleotide against insulin-like growth factor type 1 receptor (IMV-001) in biodiffusion chambers (BDC). We recently reported IGV-001-treated newly diagnosed glioblastoma patients had a median progression free survival of 38.4 months compared with 8.3 months in historical standard-of-care-treated patients (P=0.0008). We have now identified key cytokines and other clinical measures that are predictive of patient outcome via serum profiling and machine learning classification. Methodology: Cytokines in sera were assayed from the day of treatment (0d), 14, 28, 42, and 150d post-treatment. Cytokines quantified included IL-1β, IL-5, IL-6, IL-8, IL-10, IL-13, IL-15, IL-17A, IL-17F, IFNγ, and TNFα. An overall survival cut point of 21.9 months was used to dichotomize ‘good’ vs. ‘poor’ outcomes. MATLAB® Classification Learner was used to train an “optimized tree” predictive model with a sub-population of samples. The trained model was subsequently used to predict survival classification for all patients based on cytokine and clinical data inputs Results: IL-8 showed predictive potential discerning good vs. poor outcomes. Median IL-8 sera concentrations from good (n=10) vs. poor (n=15) responders showed significant differences on 28d (good=7.42 pg/ml, poor=14.73 pg/ml, p=0.027) and 42d (good=4.56 pg/ml, poor=19.87 pg/ml, p=0.0025) compared to 0d (good=8.17 pg/ml, poor=7.42 pg/ml, p=0.8). Training the classification model, a maximum of 4 splits (“Gini’s diversity index” criterion) were possible while achieving 100% correct classification of the training dataset (n=8, good=4, poor=4). Applying the trained model on the full dataset (n=25, good=10, poor=15), resulted in 100% correct classification of patient’s actual clinical outcomes. Analysis of data distributions used by the trained model indicated key cytokine were IL-8, IL-6, and IL-17A. Conclusions: A classification model has been developed to predict long-term “good” vs. short-term “poor” outcomes of IGV-001-treated patients. In particular, we identified IL-8, IL-6 and IL-17A as key immune correlates of patient outcome as early as d14 post-treatment and sustained beyond standard of care treatment. These cytokines are known to correlate with tumor burden, validating the utility of our model. Further inclusion of data from new patients from an upcoming Phase IIb (ClinicalTrials.gov Identifier: NCT04485949) will be utilized to strengthen the model’s predictive capabilities and the potential identification of patient populations most likely to benefit from IGV-001 immunotherapy. Citation Format: Christopher Uhl, Jenny Zilberberg, Charles B. Scott, David W. Andrews, Mark Exley. Machine learning algorithm identifies key serum cytokines associated with evidence of clinical activity in patients treated with personalized immunotherapeutic platform (IGV-00) [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 2782.