Functional drug response profiling predicts clinical outcomes and shows utility for treatment stratification in AML

Abstract Background Selecting the optimal therapy in acute myeloid leukemia (AML) can be clinically challenging—particularly for patients with intermediate genetic risk or borderline fitness for intensive treatment. While genomic stratification such as ELN 2022/2024 is prognostic on a population level, it does not reliably predict individual response to specific therapies. Functional precision medicine, which measures drug sensitivity directly in malignant cells of individual patients (Letai, A., Bhola, P. & Welm, A.L., Cancer Cell 2022), may offer actionable insights that complement genetic stratification. Methods We conducted an observational study (AC2305) to evaluate the association between ex vivo drug response (EVDR) to standard-of-care therapies and clinical outcomes in AML. EVDR scores were generated from diagnostic cryopreserved samples (n=91) using a microscopy-based single-cell platform. Scores for intensive (7+3, CPX-351, ±FLT3 inhibitors) and low-intensity (Venetoclax/Azacitidine, Azacitidine) regimens were calculated from dose-dependent leukemic blast killing, excluding toxicity to mature monocytes and lymphocytes. Associations between EVDR and clinical response—defined as composite marrow remission (mCR: composed of CR, CRi, CRh and MLFS)—as well as overall survival (OS), were analyzed by multivariable logistic regression and Cox regression models incorporating additional clinicopathological covariates. Results EVDR scores corresponding to the administered regimen were significantly higher in responders than in non-responders, with strong discriminatory performance (AUC=0.83). In univariate analysis, EVDR (per 0.1 unit increase) was strongly associated with mCR likelihood (OR=2.30, p=4.8×10⁻⁶), and this association remained significant in multivariate analysis (OR=1.91, p=9.7×10⁻4), alongside induction therapy intensity and ELN risk category. Similarly, EVDR was a significant predictor of improved OS, with a multivariate hazard ratio (HR) of 0.78 (p=4.8×10⁻³), whereas Kaplan–Meier analysis confirmed that patients with EVDR scores above the optimal Youden threshold had significantly longer survival (log-rank p=3.9×10⁻⁶). Importantly, these associations remained robust in post-hoc subgroup analyses, where EVDR predicted mCR and OS across both intensive and low-intensity regimens. The ELN intermediate-risk category encompasses genetically heterogeneous cases without prognostically favorable or adverse mutations, and often lacking molecular targets for therapy selection (Döhner, H. et al., Blood 2022). To address whether integrating EVDR into existing genetic stratification frameworks could improve risk assessment, we focused on ELN-intermediate patients across both intensive and low-intensity treatment arms. Using a threshold derived from the Youden index, EVDR predicted mCR with 80% accuracy (12/15 patients) in intensively treated patients and 100% accuracy (11/11 patients) in those treated with low-intensity regimens, as well as stratified OS in this subgroup. These findings suggest EVDR may provide predictive insight where ELN risk offers limited discrimination. Lastly, we focused on the potential of EVDR-based predictions to inform individualized treatment selection. Among patients who failed to achieve mCR following intensive induction therapy (n=25), we detected a subgroup of patients exhibiting a higher modeled probability for responding to the Ven/Aza therapeutic regimen (termed “switch candidates”). Their Ven/Aza modeled probabilities closely matched those of clinical Ven/Aza-treated responders, suggesting they might have benefited from an alternative regimen. We then asked whether genetic reclassification alone would have supported such a switch in early refractory settings to intensive induction therapy. Whereas all switch candidates were classified as adverse or intermediate by ELN2022, ELN2024 reclassified only 8 patients (47%) into a more favorable genetic class—indicating that genetic re-stratification alone would not have suggested a change of regimen for every switch candidate. Conclusions EVDR profiling is a strong independent predictor of remission and survival in AML and provides complementary value to ELN classification—particularly in intermediate-risk and clinically ambiguous cases. Beyond response prediction, EVDR can enable individualized modeling of treatment benefit to different therapeutic regimens and could support treatment decisions in patients where genetic guidance is limited.