Patient-tailored design of AML cell subpopulation-selective drug combinations

Abstract

AbstractThe extensive primary and secondary drug resistance in acute myeloid leukemia (AML) requires rational approaches to design personalized combinatorial treatments that exploit patient-specific therapeutic vulnerabilities to optimally target disease-driving AML cell subpopulations. However, the large number of AML-relevant drug combinations makes the testing impossible in scarce primary patient cells. This combinatorial problem is further exacerbated by the translational challenge of how to design such personalized and selective drug combinations that do not only show synergistic effect in overall AML cell killing but also result in minimal toxic side effects on non-malignant cells. To solve these challenges, we implemented a systematic computational-experimental approach for identifying potential drug combinations that have a desired synergy-efficacy-toxicity balance. Our mechanism-agnostic approach combines single-cell RNA-sequencing (scRNA-seq) with ex vivo single-agent viability testing in primary patient cells. The data integration and predictive modelling are carried out at a single-cell resolution by means of a machine learning model that makes use of compound-target interaction networks to narrow down the massive search space of potentially effective drug combinations. When applied to two diagnostic and two refractory AML patient cases, each having a different genetic background, our integrated approach predicted a number of patient-specific combinations that were shown to result not only in synergistic cancer cell inhibition but were also capable of targeting specific AML cell subpopulations that emerge in differing stages of disease pathogenesis or treatment regimens. Overall, 53% of the 59 predicted combinations were experimentally confirmed to show synergy, and 83% were non-antagonistic, as validated with viability assays, which is a significant improvement over the success rate of randomly guessing a synergistic drug combination (5%). Importantly, 67% of the predicted combinations showed low toxicity to non-malignant cells, as validated with flow-based population assays, suggesting their selective killing of AML cell populations. Our data-driven approach provides an unbiased means for systematic prioritization of patient-specific drug combinations that selectively inhibit AML cells and avoid co-inhibition of non-malignant cells, thereby increasing their likelihood for clinical translation. The approach uses only a limited number of patient primary cells, and it is widely applicable to hematological cancers that are accessible for scRNA-seq profiling and ex vivo compound testing.