We propose a novel DTA prediction method (HiSIF-DTA) that enables hierarchical fusion of protein semantic information to enrich protein representations and enhance DTA prediction performance. In order to select the optimal hyperparameters from the complex hyperparameter space for HiSIF-DTA, we conducted a 5-fold cross validation on the training set of Davis. In this approach the training set of Davis was further divided into five equal parts, with each part being used as the validation set in rotation. The models were trained on each fold independently, and the average of the results on validation set from the five models was taken to represent the overall performance of the proposed model under a specific set of parameter combinations. These files contain the predicted values of HiSIF-DTA on the trainning set of Davis under different parameter combinations.