A Machine Learning–Based Strategy Predicts Selective and Synergistic Drug Combinations for Relapsed Acute Myeloid Leukemia

Abstract

Combination therapies are one potential approach to improve the outcomes of patients with relapsed/refractory (R/R) disease. However, comprehensive testing in scarce primary patient material is hampered by the many drug combination possibilities. Furthermore, inter- and intrapatient heterogeneity necessitates personalized treatment optimization approaches that effectively exploit patient-specific vulnerabilities to selectively target both the disease- and resistance-driving cell populations. In this study, we developed a systematic combinatorial design strategy that uses machine learning to prioritize the most promising drug combinations for patients with R/R acute myeloid leukemia (AML). The predictive approach leveraged single-cell transcriptomics and single-agent response profiles measured in primary patient samples to identify targeted combinations that coinhibit treatment-resistant cancer cells individually in each sample of patients with AML. Cell type compositions evolved dynamically between the diagnostic and R/R stages uniquely in each patient, hence requiring personalized drug combination strategies to target therapy-resistant cancer cells. Cell population–specific drug combination assays demonstrated how patient-specific and disease stage–tailored combination predictions led to treatments with synergy and strong potency in R/R AML cells, whereas the same combinations elicited nonsynergistic effects in the diagnostic stage and minimal coinhibitory effects on normal cells. In preliminary experiments on clinical trial samples, the approach predicted clinical outcomes of venetoclax–azacitidine combination therapy in patients with AML. Overall, the computational–experimental approach provides a rational means to identify personalized combinatorial regimens for individual patients with AML with R/R disease that target treatment-resistant leukemic cells, thereby increasing their likelihood of clinical translation.