Machine Learning-Based Gene Signature Detection Highlights CXCL12 as a Key Marker for AML Prediction

Authors

• Selahattin Alperen Uysal, Ege UniversityFollow
• Burçin Kaymaz, Ege UniversityFollow

Abstract

Background: Acute myeloid leukemia (AML) is a severe hematologic malignancy marked by uncontrolled proliferation and impaired differentiation of myeloid cells, disrupting normal hematopoiesis and resulting in poor clinical outcomes. Conventional diagnostic approaches often lack the precision to accurately classify AML subtypes, necessitating the integration of advanced computational methods to improve diagnostic and therapeutic strategies.

Objectives: This study aimed to apply machine learning (ML) techniques to transcriptomic data in order to identify a concise and informative gene signature capable of distinguishing AML cases from normal samples. Additionally, 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 (GDC) portal. Feature selection was conducted using SelectKBest with chi-square scoring, identifying the top 10 AML-associated genes. To address class imbalance, SMOTE was employed. Several ML models—RandomForest, XGBoost, LightGBM, and a Stacking ensemble—were trained and optimized through hyperparameter tuning to classify AML versus normal samples.

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

Conclusions: This study identified a small yet informative gene set that accurately differentiates AML from normal samples, with LightGBM showing the highest predictive accuracy. The involvement of immune and inflammatory genes aligns with AML’s biological underpinnings. These findings suggest clinical potential for ML-based tools in 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.

DOI

10.56808/2673-060X.5703

First Page

1

Last Page

14

Recommended Citation

Uysal, Selahattin Alperen and Kaymaz, Burçin () "Machine Learning-Based Gene Signature Detection Highlights CXCL12 as a Key Marker for AML Prediction," Chulalongkorn Medical Journal: Vol. 70: Iss. 5, Article 2.
DOI: https://doi.org/10.56808/2673-060X.5703
Available at: https://digital.car.chula.ac.th/clmjournal/vol70/iss5/2