Deep Learning-Based Prediction of Pathogenicity for ABL1 Protein Variants Using Sequence Representation

The ABL1 gene encodes a non-receptor tyrosine kinase implicated in leukemia and other genetic disorders. This study presents a deep learning-based approach for predicting the pathogenicity of ABL1 single amino acid variants (SAVs) using amino acid sequence representations. Variant data collected from UniProt, ClinVar, dbSNP, and EVS databases were encoded through embedding and transformer-based methods. Multiple architectures-Feedforward, Conv1D, BiLSTM, Transformer, and Prot-BERT-were evaluated under six data partitioning scenarios. The Convolutional Neural Network achieved the highest performance (AUC = 0.86; Accuracy = 93%; Specificity = 0.93) in distinguishing benign from pathogenic variants. The findings demonstrate the potential of sequence-centered deep learning frameworks for accurate variant effect prediction and support the integration of AI-assisted tools in computational genomics.