Modeling progression and treatment of AML using the Thales QSP software platform.

Abstract

e18506 Background: Acute myeloid leukemia (AML) is an aggressive hematological malignancy characterized by the abnormal maturation and proliferation of myeloblasts in the bone marrow. Clinical need exists for therapies which are effective against AML, including for common genomic aberrations such as NPM1, FLT3, or KMT2A. Here, a QSP platform model of newly diagnosed (ND) and relapsed/refractory (RR) AML was developed, and a virtual population of AML patients was created by training and validating to clinically relevant measures from clinical trials. This model and virtual population can be used to evaluate new therapies for treating AML patient populations, including mutational subpopulations of interest. Methods: A QSP model of AML was created in Thales, a model building platform that automates and streamlines the QSP modeling process. The model structure includes representations of hematopoietic stem cells and mature immune cells, immune cell killing of cancer cells, and cell proliferation, differentiation, and migration. Additionally, each virtual patient is capable of having a combination of genomic aberrations common to AML (NPM1, FLT3, and/or KMT2A) which affect disease progression and treatment response. Parameter values were chosen to align with existing published AML models or were estimated from published in vitro or in vivo data. A virtual population of 404 AML patients was generated by simultaneously fitting and validating to publicly available clinical data from 23 different trials (ND and RR), spanning 13 unique therapies administered either alone or in combinations. Key clinical measures were derived from model predictions of blast, neutrophil, and platelet counts in the bone marrow and blood and were used in training. These included ELN endpoints (CR, CRi, CRh, PR, ORR), duration of response, relapse, and minimum residual disease. Results: The simulations closely match the training data, with >72% of clinical endpoint data falling within the model’s 90% confidence intervals and >78% for the validation data. The results capture several important dynamics in treatment of AML including differential responses in ND and RR populations, as well as stratification of treatment efficacy by genomic aberration (e.g. observing worse outcomes in FLT3+ individuals than in FLT3-). Conclusions: These results provide proof-of-concept that AML progression and treatment is able to be modeled with a QSP platform model. The model and virtual population accurately predict endpoint data for many clinical trials and therapeutic regimens. Clinically relevant predictions for novel compounds such as optimal first-in-human dose, special populations of interest, or novel therapeutic combinations can be made using the trained model and virtual population. Furthermore, proprietary data from individual patients or compound assays could be included to further enhance the model’s predictions.