Drug-Target Binding Affinity Prediction: Evaluating the Impact of Dataset Size on Model Performance

Abstract

Drug-target interaction (DTI) prediction plays a crucial role in modern drug discovery. However, real-world DTI prediction is often challenged due to limited availability of known interaction data on drugs and targets. This study investigates how the size of a training dataset influences the performance of a machine learning model, in particular, Kronecker Regularized Least Square model evaluated using Concordance Index under four experimental settings that represent the realistic scenarios in DTI prediction. These four settings are based on whether the test set contains entirely novel drugs and targets, share either drugs or targets in the training set or includes drugs and targets that are both present during the training phase. The results of the experiment revealed that the model performance heavily varies depending on the experimental setting. Although the model showed improvement of prediction performance under each setting with increased training data size, some settings showed much improvement and better generalization ability than the others. As expected, the model performed best when the model has seen interactions involving the drugs and targets in the test set during training. The model performed well when predicting interactions for unseen targets but struggled to generalize well with novel drugs or both novel drugs and targets. The findings of the study highlight the importance of addressing each experimental setting individually. While some experimental settings benefited significantly well with the available training data size, others showed only slight improvement though with an upward trend suggesting that more data may help to further improve the performance.