A Transfer Learning-Based Approach to Unveil Kinase-Specific Phosphorylation Sites of Understudied Kinases

Phosphorylation plays a pivotal role in regulating protein activities; yet, it is still time-consuming, costly, and human-error prone to identify specific phosphorylation sites for kinases. Existing studies have proposed kinase-specific phosphorylation site prediction tools, primarily focusing on kinases with a sufficient number of experimentally validated phosphorylation sites. Nevertheless, the majority of kinases lack such experimental data or have unknown targets. This study developed and validated a novel transfer learning-driven approach to predict specific phosphorylation sites for understudied kinases. By building upon pre-trained models at the corresponding kinase group levels, custom models have been created by fine-tuning them on layers of closely related kinases. Importantly, the encoding of substrate-specific site locations serves as critical features in the predictive model. The proposed approach achieves a balanced accuracy of 78.3%, 81.9%, and 83.7% for 18, 17, and 13 understudied kinases respectively in the Other, STE, and TKL groups.