Discovery of Marker Genes in Adult T Cell Leukemia (ATL) Pathogenesis with Machine Learning Models and Performance Comparison

Hematologic cancers are often diagnosed after symptoms become apparent, which can make it difficult to control the disease and implement effective treatment strategies. Studying gene expression profiles is vital for early diagnosis and the development of treatment strategies for hematologic cancers such as T-cell leukemia. The motivation of this study is to reveal the molecular mechanisms in the pathogenesis of this disease by comparing the whole gene expression profile in Adult T-cell Leukemia (ATL) cells and CD4+T cells of healthy individuals. For this aim, several machine learning algorithms, Naive Bayes, K-Nearest Neighbor, Support Vector Machine, Random Forest, C4.5, Logistic Regression, Linear Discriminant Analysis and Artificial Neural Network algorithms were used. Their performance was compared on the GSE33615 dataset by using 5-fold cross validation with stratified sampling. Among these, Artificial Neural Network stood out with an AUC of 0.98 and an F1 score of 0.93. It was followed by SVM with an AUC of 0.97 and 0.957 F1 score. In addition to performance comparison, information gain ratio, SHAPLEY metric and correlation values were calculated for the detection of genes causing ATL. Among the models, the three with the highest performance (ANN, SVM, RF) were selected, and the top ten most significant genes were identified for each. Considering the intersection of these gene sets, ZSCAN18, PLK3, and NELL2 were found to be associated with the related disease. These genes may contribute to Adult T-cell Leukemia pathogenesis through their roles in cell cycle regulation, transcriptional control, and oncogenic signaling. Further investigation is needed to clarify their precise molecular mechanisms in the related disease.