Abstract 7425: AttentionAML: An attention-based deep learning model for accurate identification of childhood acute myeloid leukemia subtypes

Abstract Acute myeloid leukemia (AML) is a rapidly progressing hematopoietic malignancy characterized by aberrant clonal expansion of abnormal myeloid progenitor cells, typically leading to bone marrow failure and compromised hematopoiesis. Pediatric AML encompasses over 20 molecular subtypes defined by genetic alterations, such as chromosomal alteration, gene fusions, mutations, or tandem duplications. Each subtype exhibits a specific response to treatment and prognosis, underscoring the critical need of accurately identifying subtypes for effective clinical management and tailored therapeutic approaches. Traditional wet lab approaches to identify AML subtypes primarily rely on morphological analysis, cytogenetic analysis, immunophenotyping, or molecular profiling, which can be labor-intensive, costly, and technically challenging for clinical practice applications. To address these challenges, we propose AttentionAML (Attention-based Multilayer Perceptron Model for Identifying Acute Myeloid Leukemia Subtypes), a novel deep learning approach that leverages the power of attention mechanisms and multilayer perceptron (MLP) to accurately classify AML subtypes based solely on transcriptomic profiling. Compared with standard MLP models, the attention-based MLP model can capture both global and local information within high-dimensional gene expression data by focusing on the most relevant features, enhancing the performance of the model in AML subtyping. Our model is trained on different cohorts of 1, 656 AML patients with known subtypes, utilizing RNA-seq data from publicly available databases such as Therapeutically Applicable Research to Generate Effective Treatments (TARGET) initiative and the St. Jude Cloud Genomics Platform. To evaluate the performance and generalizability of AttentionAML, we conducted rigorous 100 times five-fold cross-validation tests. Our results suggested that AttentionAML consistently outperformed both conventional machine learning models (e.g., ensemble SVM) and standard deep learning models (e.g., MLP model), with remarkable performance across different metrics, including the accuracy of 97.15%, precision of 97.13%, recall of 97.15%, F1 score of 97.02%, and the Matthews correlation coefficient (MCC) of 0.9674. In addition, by leveraging the hidden layer embedding information, AttentionAML gave clearer and more insightful visualization of the global landscape of all AML patients compared to tSNE, providing better separation of AML subtypes and preserving both local and global data structures. We believe that AttentionAML represents a significant progression in characterizing pediatric AML, which can potentially improve downstream risk stratification, optimize treatment strategies, and advance precision medicine for AML as well as other pediatric, adolescent, and young adults (AYA). Citation Format: Lusheng Li, Joseph D. Khoury, Jieqiong Wang, Shibiao Wan. AttentionAML: An attention-based deep learning model for accurate identification of childhood acute myeloid leukemia subtypes [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7425.