Identification of methylation signatures associated with CAR T cell in B-cell acute lymphoblastic leukemia and non-hodgkin's lymphoma

Abstract

CD19-targeted CAR T cell immunotherapy has exceptional efficacy for the treatment of B-cell malignancies. B-cell acute lymphocytic leukemia and non-Hodgkin's lymphoma are two common B-cell malignancies with high recurrence rate and are refractory to cure. Although CAR T-cell immunotherapy overcomes the limitations of conventional treatments for such malignancies, failure of treatment and tumor recurrence remain common. In this study, we searched for important methylation signatures to differentiate CAR-transduced and untransduced T cells from patients with acute lymphoblastic leukemia and non-Hodgkin's lymphoma. First, we used three feature ranking methods, namely, Monte Carlo feature selection, light gradient boosting machine, and least absolute shrinkage and selection operator, to rank all methylation features in order of their importance. Then, the incremental feature selection method was adopted to construct efficient classifiers and filter the optimal feature subsets. Some important methylated genes, namely, SERPINB6, ANK1, PDCD5, DAPK2, and DNAJB6, were identified. Furthermore, the classification rules for distinguishing different classes were established, which can precisely describe the role of methylation features in the classification. Overall, we applied advanced machine learning approaches to the high-throughput data, investigating the mechanism of CAR T cells to establish the theoretical foundation for modifying CAR T cells.

Keywords: B-cell acute lymphocytic leukemia; B-cell acute non-Hodgkin’s lymphoma; CAR T cell; classification algorithm; classification rule; feature selection.

Figure 1

Flow chart of the entire analytical process. The 865,859 methylation probe signals on patients with B-cell malignancies were ranked according to feature importance by using three feature ranking algorithms, namely, MCFS, LightGBM, and LASSO. Then, three ordered feature lists were fed into the incremental feature selection (IFS) method, which incorporates four classification algorithms. Finally, based on the IFS results, the essential genes, efficient classification models and classification rules were extracted.

Figure 2

IFS curves for displaying the performance of four classification algorithms on the feature list yielded by MCFS method. The best classifiers on different algorithms yield the weight F1 values of 0.861, 0.864, 0.912 and 0.827, respectively, which use top 591, 680, 354 and 952, respectively, features in the list.

Figure 3

Performance of the best classifiers using different classification algorithms and feature lists on four classes. (A) Feature list generated by MCFS method; (B) Feature list generated by LightGBM method; (C) Feature list generated by LASSO method.

Figure 4

IFS curves for displaying the performance of four classification algorithms on the feature list yielded by LighGBM method. The best classifiers on different algorithms yield the weight F1 values of 0.956, 0.938, 0.975 and 0.950, respectively, which use top 181, 12, 140 and 43, respectively, features in the list.

Figure 5

IFS curves for displaying the performance of four classification algorithms on the feature list yielded by LASSO method. The best classifiers on different algorithms yield the weight F1 values of 0.912, 0.943, 0.987 and 0.987, respectively, which use top 9, 12, 28 and 111, respectively, features in the list.

Figure 6

Venn diagram to show the intersection of the optimal gene sets for MCFS, LightGBM, and LASSO. Three genes are contained in two optimal gene sets, indicating their importance.

Figure 7

The number of rules extracted from the decision tree built on feature lists yielded by MCFS, LightGBM, and LASSO, respectively, on four classes.