Evaluating the Transferability of Deep Learning-Based Drug-Kinase Interaction Prediction Models: 15th International Conference on Bioinformatics and Biomedical Technology, ICBBT 2023

Aberrant kinase activation is involved in numerous disease processes and has become an important therapeutic target for diseases such as tumors and inflammation. Therefore, the development of kinase inhibitors is crucial in medical treatment. Drug-target interaction (DTI) prediction has gained increasing interest in the field of drug design due to its time-saving and cost-effective advantages. Previous studies have proposed several statistical methods and machine learning models to estimate the binding affinity of drugs to targets, with graph convolutional neural network (GCN) learning methods being one of the most advanced. However, the performance of deep learning methods heavily relies on the diversity of the training data. As a result, they often exhibit attenuated generalization ability especially when searching for drug candidates for new targets with limited or no known DTI information. Therefore, this study aims to evaluate the transferability of GCN-based DTI models. Specifically, DTI models were developed using the entire collection of inhibitory molecules towards all known human kinases and subsets towards four kinase families (TK, TKL, AGC, CMGC). The aim was to test whether the models are sensitive to differences in data from different kinase families and whether models trained on some kinase family/ies can be used for predicting new DTIs on other families. The results indicate that equivalent or better predictions could be achieved using a smaller training set of the same family compared to using the whole human kinome data. However, the prediction accuracy decreases slightly when the model uses different kinase family data (e.g., using TK to predict AGC). Possible reasons for the poor performance of the model on external testing data were analyzed, and solutions for future studies were proposed to enhance the model's performance.