Abstract
The Cystic Fibrosis Transmembrane Conductance Regulator (CFTR) protein, crucial for chloride ion transport across epithelial cells, requires ATP binding for proper function. Mutations within CFTR's nucleotide-binding domains (NBDs) disrupt this process, leading to cystic fibrosis. Accurately predicting ATP binding sites within CFTR is essential for understanding its function and developing targeted therapies. However, the unique nature of CFTR as an ion channel within the ATP-binding cassette (ABC) transporter family poses challenges for general prediction methods. We present CFTR_TL, a novel approach that leverages transfer learning to enhance ATP binding site prediction in CFTR. Our approach involves training a base model on a broad data set of ATP-binding proteins, followed by fine-tuning with a data set enriched in ABC transporters, capitalizing on their functional similarity to CFTR. By utilizing a multiwindow convolutional neural network (CNN) to effectively capture spatial patterns, CFTR_TL achieves superior performance compared to traditional prediction methods. The resulting model demonstrates improved accuracy and specificity in identifying critical binding residues within CFTR. This approach not only provides a powerful tool for CFTR research but also offers a generalizable framework for tailoring prediction models in other protein families.