Abstract 4342: Predictive precision medicine platform accurately predicts individual patient response to AML treatments to maximize outcomes

Abstract

Abstract Introduction: Offering the optimal frontline treatment to a patient with acute myeloid leukemia (AML) requires trading off expected benefit and risk. Typical standard of care intensive induction chemotherapy (e.g., cytarabine plus idarubicin (7+3)) results in high clinical response rates. However, many patients receive a less intensive regimen (e.g., venetoclax plus decitabine (VenDec)) because their individual toxicity risk is high based on lack of medical fitness. Predicting an individual patient’s clinical response prior to treatment has the potential to increase the benefit/risk ratio (therapeutic index) for some patients and optimize their treatment selection. Here, we demonstrate the ability of an automated high-throughput, multi-color flow cytometry predictive precision medicine platform (PPMP) to predict response to 7+3 or VenDec. Methods: To assess correlation between PPMP-predicted and actual clinical response to 7+3 or VenDec in clinical trial NCT04263181, pre-induction blood samples were collected from 31 patients of which 18 received 7+3 (all newly diagnosed (ND) AML) and 13 VenDec (7 ND AML, 5 secondary AML, 1 MDS). We measured drug effects on leukemic blasts by applying a cutoff at the total blast count that optimizes separation between predicted responders and non-responders (“conventional approach”) or by a machine learning (ML) approach considering multiple cell populations. For the former approach, training sets represented 13 patients for 7+3 and 8 for VenDec. Both 7+3 and VenDec models were validated with 5 patients. For the ML approach, the model was trained on all 13 VenDec patients and monitored using leave-one-out cross-validation. Results: For the conventional approach, predicted and true clinical responses were highly correlated for 7+3 (AUC = 0.91) and VenDec (AUC = 0.81), with 100% precision (positive predictive values (PPV)) for both, i.e., all predicted responders were true clinical responders. Some true clinical responders were not identified (negative predictive value (NPV) = 67% for 7+3 and 57% for VenDec), resulting in an accuracy of 94% (7+3) and 77% (VenDec). To maximize NPV and accuracy for predicting VenDec clinical outcomes, we applied a novel ML-based algorithm to integrate the behavior of malignant and non-malignant cell populations, yielding a model with 100% accuracy (100% PPV and NPV). Additional outcome data, including overall survival, are under evaluation. Summary: Total blast-based predictions yielded accuracies of 94% for 7+3 and 77% for VenDec. An ML algorithm for VenDec considering additional cell populations increased the accuracy to 100%. Further studies will expand patient numbers. We plan to use this platform to inform our frontline decision making with the goal to maximize the therapeutic benefit/risk ratio and ensure that the most appropriate frontline therapy is used for each individual patient. Citation Format: Meagan A. Jacoby, John S. Welch, Peter Westervelt, Matthew Christopher, Geoffrey L. Uy, Ravi Vij, Keith E. Stockerl-Goldstein, Brad S. Kahl, Iskra Pusic, John F. DiPersio, Mark A. Schroeder, Miriam Y. Kim, Todd A. Fehniger, Armin Ghobadi, Christine J. Gu, Wade Anderson, Kathryn Vanderlaag, Kamran Ali, Camille Pataki, Markus D. Lacher. Predictive precision medicine platform accurately predicts individual patient response to AML treatments to maximize outcomes. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 4342.