Improved Prediction of Venetoclax and Azacitidine Efficacy in Acute Myeloid Leukemia through Machine Learning [AML patients]

Despite the transformative impact of venetoclax-azacitidine in treating acute myeloid leukemia (AML), reliable markers for accurately predicting patient responses remain urgently needed. To address this challenge, we employed a multidisciplinary approach that combined transcriptomic profiling, ex vivo drug sensitivity testing, functional assays, and clinical data to identify robust predictors of venetoclax-azacitidine response. We pinpointed a set of core genes linked to both ex vivo and in vivo drug responsiveness, validated through CRISPR-Cas9 screens in the setting of both venetoclax and azacitidine therapies. In particular, silencing BCL2L1 and PINK1 preferentially enhanced response to the venetoclax-azacitidine treatment. Building on these insights, we further developed and validated an 8-gene random forest model (RF8) that demonstrated high specificity and sensitivity in four independent cohorts comprising 498 patients. This model was capable of distilling downstream effects of genetic alterations to assist in predicting treatment response and outperformed existing genetic mutation-based signatures. Furthermore, the RF8 score demonstrated a nearly monotonic relationship with venetoclax-azacitidine response probabilities and patient outcomes, enabling precise stratification of patients. These findings illustrated the feasibility of translating integrated transcriptomic and drug-response profiling data into more refined risk stratification approaches, offering a new avenue for optimizing clinical decision-making in AML.