Search

Introduction: Ex vivo drug sensitivity studies of samples derived from acute myeloid leukemia (AML) patients have been shown to be predictive of in vivo response. These findings are based on a limited number of well-characterized agents for which in vivo patient response data and ex vivo drug sensitivity data—on that same patient—are available. To show the feasibility of scaling such ex vivo studies to large drug screens, we characterized the reproducibility of expression-based models of drug response across two independent data sets—one generated at the Oregon Health and Science University (OHSU) and the second at the Institute for Molecular Medicine Finland (FIMM). Methods: We harmonized two large-scale AML ex vivo studies screened for drug response and profiled transcriptomically—OHSU (303 AML patient samples and 160 drugs) and FIMM (48 AML samples and 480 drugs). The two panels have 94 drugs in common. Log-logistic curves were fit to the dose-response data and area under the dose-response curves (AUCs) were calculated. Predictive modeling using Ridge regression or an integrative Bayesian approach was performed for each drug AUC independently using 202 highly-variable and/or cancer-associated genes as features. Results: For each of the 94 drugs in common between the two data sets, we trained a Ridge regression model on the OHSU data set, used the model to predict response in the FIMM data set, and calculated the Pearson correlation between the predicted and observed FIMM responses. 41 of the 94 drug models had a positive and statistically significant correlation [false discovery rate (FDR) < 20%; mean ρ = 0.43; 95% CI = 0.29 – 0.77]. Drugs corresponding to the top decile of these significant models (mean ρ = 0.54; 95% CI = 0.48 – 0.77) clustered into four primary classes: MEK inhibitors (PD184352, Selumetinib, and Trametinib), EGFR/VEGFR inhibitors (Cabozantinib, Erlotinib, Foretinib, and Sorafenib), and singletons Venetoclax and Sirolimus. To confirm these results, we applied a second modeling approach—an integrative Bayesian machine learning method—that allows systematic combination of both data sets. Training and evaluation of this approach using 10-fold cross validation yielded 82 positive and statistically significant correlations (FDR < 20%; mean ρ = 0.35; 95% CI = 0.13 – 0.58). Five of nine drugs (Cabozantinib, Selumetinib, Sirolimus, Sorafenib, and Trametinib) corresponding to the top decile of these significant models (mean ρ = 0.54; 95% CI = 0.49 – 0.60) overlapped with drugs from the top decile of Ridge results (one-sided Fisher p = 2.5 x 10-4) Conclusions: Our results using independent data sets and two statistical approaches suggest that certain drugs (including MEK and EGFR/VEGFR inhibitors) are amenable to expression-based predictive modeling in AML. Future work will focus on inferring individual biomarkers of response. Citation Format: Brian S. White, Suleiman A. Khan, Muhammad Ammad-ud-din, Swapnil Potdar, Mike J. Mason, Cristina E. Tognon, Brian J. Druker, Caroline A. Heckman, Olli P. Kallioniemi, Stephen E. Kurtz, Kimmo Porkka, Jeffrey W. Tyner, Tero Aittokallio, Krister Wennerberg, Justin Guinney. Gene expression predicts ex vivo drug sensitivity in acute myeloid leukemia [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 3883.