Identifying AML patients most likely to benefit from venetoclax using apoptotic gene ratios

Abstract Venetoclax is widely used to treat patients with acute myeloid leukemia (AML) who are older or those who are not candidates for intensive chemotherapy. But some patients don't respond well, and it's still hard to predict who will benefit from it. Lab tests that measure how primed leukemia cells are for apoptosis can help, but they're not easy to use in clinical practice. Therefore, in this study, we looked into an alternative approach using RNA-based markers, specifically the gene expression ratios related to apoptosis to predict response to venetoclax and the patient outcomes, regardless of age or mutation status. We analyzed 309 AML patients' data from the BeatAML2 dataset with available RNA-sequencing, mutational, and ex vivo drug sensitivity data. We calculated four gene expression ratios: BCL2/MCL1, BCL2L11/PMAIP1, MCL1/BAX, and BCL2L1/BAX. Venetoclax sensitivity was measured as the area under the dose-response curve (AUC), with lower AUC values indicating greater sensitivity. We used correlation and regression analysis to test associations between gene ratios and AUC. We also used XGBoost models to predict complete remission (CR) and 2-year overall survival (OS), incorporating age group, common mutations (TP53, FLT3, NPM1, RUNX1, ASXL1), and the gene expression ratios. The interpretation of the model was performed using SHAP. The BCL2/MCL1 and BCL2L11/PMAIP1 gene expression ratios were both strongly linked to how well patients responded to venetoclax with spearman correlation coefficients (ρ) of 0.51 and -0.52 respectively (p<0.005). This pattern was seen across all age groups (18-30 (n=28), 30-45 (n=44), 45-60 (n=64), and 60+ (n=173)). For BCL2/MCL1, ρ values ranged from 0.36 to 0.55, and for BCL2L11/PMAIP1, the values ranged from -0.40 to -0.54, across patient age groups. For BCL2L11/PMAIP1, ρ ranged from 0.40 to -0.54. In multivariable linear regression too, the BCL2/MCL1 ratio was a significant predictor of venetoclax response. We found the regression coefficient (β) to be -206 (p<0.005), which means that for each 1-unit increase in the BCL2/MCL1 ratio, the AUC would decrease by 206 units, demonstrating improved drug sensitivity. However, no interaction was observed between BCL2/MCL1 and age groups. Likewise, including expression ratios improved the XGBoost model's ability to predict clinical outcomes, with AUCs of 0.715 for CR and 0.729 for 2-year OS. SHAP analysis confirmed that BCL2/MCL1 and BCL2L11/PMAIP1 were the most important predictive features, exceeding the contribution of all five tested mutations. Apoptotic gene expression ratios, BCL2/MCL1 and BCL2L11/PMAIP1 were strong predictors of venetoclax sensitivity and clinical outcomes, regardless of age or mutation status. Overall, we demonstrate that RNA-based apoptotic gene expression ratios can serve as a practical alternative to lab tests that measure apoptotic priming, which can be difficult to implement in routine clinical practice. By using information already captured in standard RNA sequencing, our approach could help clinicians better select patients likely to benefit from venetoclax and improve treatment outcomes in AML.