Machine learning-based gene signature detection highlights CXCL12 as a key marker for acute myeloid leukemia prediction

Abstract

Background: Acute myeloid leukemia (AML) is a severe hematologic malignancy that is marked by the uncontrolled proliferation and impaired differentiation of myeloid cells, which disrupts normal hematopoiesis and results in poor clinical outcomes. Conventional diagnostic approaches often lack the precision needed to accurately classify AML subtypes; therefore, integrating advanced computational methods is required to improve diagnostic and therapeutic strategies.

Objectives: This study aimed to apply machine learning (ML) techniques to transcriptomic data to identify a concise, informative gene signature that can distinguish AML cases from normal samples. In addition, the study sought to evaluate the performance of predictive models in supporting AML prediction and facilitating biomarker discovery.

Methods: Gene expression data were obtained from the TARGET-AML project via the Genomic Data Commons portal. Feature selection was performed using SelectKBest with chi-square scoring to identify the top 10 AML-associated genes. To address class imbalance, SMOTE was employed. Several ML models, including Random Forest, XGBoost, LightGBM, and a Stacking ensemble, were trained and optimized via hyperparameter tuning to classify AML vs. normal samples.

Results: The selected gene panel included CXCL12, SELENBP1, SLC4A1, IFIT1B, ALAS2, CAMP, OLFM4, DEFA3, CRISP3, and HBG1, which are functionally associated with immune response, inflammation, and hematopoietic processes. All models demonstrated robust classification performance: Random Forest (88.6%), XGBoost (88.8%), LightGBM (89.4%), and Stacking (89.4%). SMOTE effectively enhanced the model performance, particularly for underrepresented classes, thereby improving precision and recall.

Conclusion: This study identified a small, yet informative gene set that accurately differentiates AML from normal samples, with LightGBM revealing the highest predictive accuracy. The involvement of immune and inflammatory genes aligns with AML’s biological foundations. These findings suggest the clinical potential of ML-based tools for AML diagnostics. Future validation on independent cohorts and the integration of multi-omic data could further improve the model’s robustness and facilitate broader applications in precision medicine for hematologic malignancies.