Abstract 180: Identifying signatures of vulnerability through machine learning in an umbrella trial for glioblastoma

Glioblastoma is characterized by intra- and inter-tumoral heterogeneity. An umbrella trial tests multiple investigational treatment arms depending on corresponding biomarker signatures. A contingency of an efficient umbrella trial is a suite of preferably orthogonal molecular biomarkers to classify patients into the likely-most-beneficial arm. Assigning thresholds of molecular signatures to classify a patient as a “most-likely responder” for one specific treatment arm is a crucial task. Gene Set Variation Analysis (GSVA) of specimens from a GBM clinical trial of methoxyamine associated differential enrichment in DNA repair pathways activities with patient response. However, the large number of DNA-repair related pathways confound confident “high” enrichment of responder, as well as obscuring to what degree each feature contributes to the likelihood of a patient9s response. Here, we utilized semi-supervised machine learning, Entropy-Regularized Logistic Regression (ERLR) to predict vulnerability classification. By first training all available data using semi-supervised algorithms we transformed unclassified TCGA GBM samples with highest certainty of predicted response into a self-labeled dataset. In this case, we developed a predictive model which has a larger sample size and potential better performance. Our umbrella trial design currently includes three treatment arms for GBM patients: arsenic trioxide, methoxyamine, and pevonedistat. Each treatment arm manifests its own signature developed by the above (or similar) machine learning pipeline based on selected gene mutation status and whole transcriptome data. In order to increase the robustness and scalability (with future more treatment arms), we also developed a multi-label classification ensemble model that9s capable of predicting a probability of “fitness” of each novel therapeutic agent for each patient. By expansion to three, independent treatment arms within a single umbrella trial, a “mock” stratification of TCGA GBM patients labeled 56% of all cases into at least one “high likelihood of response” arm. Predicted vulnerability using genomic data from preclinical PDX models placed 4 out of 6 models into a “high likelihood of response” regimen. Our utilization of multiple vulnerability signatures in an umbrella trial demonstrates how a precision medicine model can support an efficient clinical trial for heterogeneous diseases such as GBM. Citation Format: Matthew Eric Lee, Nanyun Tang, Manmeet Ahluwalia, Ekokobe Fonkem, Karen Fink, Harshil Dhurv, Harshil Dhurv, Michael E. Berens, Sen Peng. Identifying signatures of vulnerability through machine learning in an umbrella trial for glioblastoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 180.