Integrating Gene Expression With Recurrent Mutations Improves Age‐Stratified Risk Prediction in Acute Myeloid Leukemia

TL;DR

Adding gene expression data to known mutations improves predicting outcomes in AML, especially for older patients.

Contribution

Combining gene expression with mutations enhances age-stratified risk prediction in AML.

Findings

Adding gene expression improved survival and remission prediction models for AML patients.

TP53 was the strongest adverse marker for survival, while NPM1 improved remission chances.

CHEK2 and CCNG1 were key genes influencing outcomes, particularly in older patients.

Abstract

Older adults with acute myeloid leukemia (AML) have inferior outcomes, yet current genetic risk models do not explicitly account for how age modifies the prognostic impact of molecular features. We hypothesized that integrating apoptosis and p53‐related gene expression with recurrent mutations would improve prediction of complete remission (CR) and 2‐year overall survival (OS), particularly across age groups.

Using the BeatAML2 dataset (805 patients; 942 specimens), we built two cohorts: a clinical cohort of 916 adults with full data and an expression‐linked cohort of 852 with matched RNA‐seq. Patients were divided into four age groups 18–30, 30–45, 45–60, and 60+ years. We tested whether adding expression of 12 apoptosis and p53‐related genes to five well‐known mutations, that is TP53, NPM1, FLT3, RUNX1, and ASXL1, could improve the prediction of CR and 2‐year OS.

Adding gene expression improved predictive performance across models. For 2‐year OS, AUCs rose from 0.765 to 0.772 in XGBoost, 0.703 to 0.843 in Random Forest, and 0.697 to 0.721 in Logistic Regression. For CR, performance improved from 0.770 to 0.851 in XGBoost, 0.811 to 0.861 in Random Forest, and 0.731 to 0.696 in Logistic Regression. Calibration was strongest for tree‐based models, and reclassification improved most with XGBoost. Multivariable regression confirmed TP53 as the most adverse marker for OS (HR: 3.07), with added risk from ASXL1 (HR: 1.53) and FLT3 (HR: 1.39). NPM1 increased the chance of remission (OR: 2.47) but did not extend survival. SHAP analysis showed that age remained the leading predictor of OS. Among genes, CHEK2 expression was most important for survival, especially in patients 60 years and older, while CCNG1 expression best predicted remission, along with BAX and MCL1.

These results demonstrate that combining gene expression with recurrent mutations makes risk prediction more accurate, especially in older patients who formed the largest group and had the poorest outcomes. Although treatment variables were not included and analysis focused on selected genes, these findings support incorporation of expression‐based features into genetic risk models and warrant prospective validation.

The authors have confirmed clinical trial registration is not needed for this submission

Figures