Integrating dual convolutional networks and BiLSTM for precision prediction of chronic myeloid leukemia from protein sequences.

Introduction: Chronic Myeloid Leukemia (CML) is a hematologic malignancy characterized by the occurrence of the Philadelphia chromosome [t(9; 22)(q34; q11)], leading to the creation of the BCR-ABL fusion gene. The fusion gene expresses a constitutively active tyrosine kinase that stimulates the uncontrolled growth and survival of myeloid cells, both a diagnostic marker and therapeutic target. Conventional diagnostic techniques, including cytogenetic examination, fluorescence in situ hybridization (FISH), and polymerase chain reaction (PCR), while accurate, remain invasive, require enormous resources, and often detect the disease at more progressed stages. Computational methods based on protein sequence analysis offer a non-invasive, scalable solution; meanwhile, contemporary machine learning methods are strongly dependent on manually designed features, limiting their ability to effectively capture long-range dependencies and subtle contextual interactions. Methods: To counter these shortcomings, we propose a Dual Convolutional Neural Network-Bidirectional Long Short-Term Memory (Dual CNN-BiLSTM) framework for the accurate prediction of CML from protein sequences. The model includes two parallel CNN modules of different kernel sizes for multi-scale motif discovery, followed by a BiLSTM layer for modeling bidirectional sequential dependencies. The combination of features is realized by concatenating ProtBERT embeddings with Pseudo Amino Acid Composition (PseAAC) and Dipeptide Composition (DPC). Results: An experimental evaluation over curated UniProtKB sets of CML-associated proteins indicates improved performance, with an accuracy of 97.5% and a 0.98 ROC-AUC. Discussion: The proposed framework delivers breakthroughs to computational oncology and enables early, non-invasive screening for CML.