Predicting Response to FLT3 Inhibitors Using Genomics-Based Computational Bio-Simulation Platform

Abstract

Background Despite rapid advances in targeted therapies, Acute Myeloid Leukemia (AML) is still suffering from drug resistance and relapse, leading to poor outcomes. That is partly because the traditional "one drug, one target” paradigm lacks consideration of the molecular complexity and heterogeneity in AML. Thus a comprehensive modeling system, integrating all known-effect mutant/drug mechanisms, is needed, which is more aligned to the characteristics of AML, providing an improved prediction accuracy of drug response. Methods We develop a computational bio-simulation platform to predict drug response based on genomics data. In brief, a static cellular signaling network is extracted from a public signaling pathway knowledge base (Reactome), then it is transformed into a computational graph based on the biochemistry knowledge in PubMed. That is used to represent a normal hematic progenitor cell (WT Digital Cell). After applying the patient specimen genomic changes to WT Digital Cell, a patient-specific disease cell is built (Disease Digital Cell). Then drug response simulation is executed based on the Disease Digital Cell. The phenotype reverse degree of Disease Digital Cell determines the final drug response score. A patient drug response dataset from the Beat AML project, is selected from ex vivo drug sensitivity assays of FLT3 inhibitors, which is used to validate the prediction performance of this computational bio-simulation platform. After generating patient-specific Disease Digital Cell based on WES sequencing data, drugs response simulation of each FLT3 inhibitor is executed. Finally, the computational bio-simulation result is compared to the BEAT AML ex vivo sensitivity assays. Results For all 533 selected sample-drug pairs the computational bio-simulation platform achieves an 84.4% accuracy with 76.1% sensitivity, 94.0% specificity, 93.5% PPV and 77.5% NPV, respectively. When stratified by risk classification (ELN2017), the adverse group (n=196) has the best performance with an accuracy of 91.3%. As for different FLT3 inhibitors, gilteritinib, which is indicated for the treatment of adult patients who have relapsed or refractory AML with FLT3 mutations, has the highest 89.3% accuracy with 81.3% sensitivity, 100.0% specificity, 100.0% PPV and 80.0% NPV. Detailed analysis shows that the model predicts correctly for cases carrying FLT3 inhibitors resistant mutations (i.e. FLT3-TKD, NRAS, PTPN11, KIT, etc.) and the inconsistent cases are mostly due to the lack of drug target mutations. Conclusions We develop a genomics-based computational bio-simulation platform for drug response prediction and validate it in the FLT3 inhibitors dataset from the Beat AML project. The comprehensive results show the computational bio-simulation platform can achieve good consistency with ex vivo sensitivity assays in drug response. Furthermore, it may contribute to choosing personalized precision therapies for AML patients. Figure 1View largeDownload PPTFigure 1View largeDownload PPT Close modal