| dc.description.abstract | Identifying drug-target interactions plays an important role in discovering drugs. Identifying, finding, and preparing drug molecule targets is the key for modern drug discovery. However, potential drug-target interactions are usually determined experimental approaches (in vivo and in vitro). Experimental approaches are expensive, require a lot of manpower, and the data are complex, making it difficult to use these methods effectively. For these reasons, the importance of simulation-based methods (in-silico) has increased and computational methods have started to be used more actively. In addition, more computational methods need to be developed to validate the interactions between drugs and their targets. In order to predict and validate the interactions between drugs and their targets by computational methods, both drugs and targets need to be mapped and to be classified with artificial intelligence techniques. As it is known, targets consist of proteins and protein sequences consist of letters. Furthermore, drug compounds are expressed in molecular codes. It is not possible to determine the interactions between drugs and their targets by computational methods without any pre-processing (mapping). The performance of the DTI (Drug-Target Interaction) prediction process varies according to the protein mapping and artificial intelligence approaches selected thus, it is important to choose the right methods in such applications. There are a number of protein mapping techniques and artificial intelligence algorithms in the literature. In this study, prediction of drug-target interactions carried out for COVID-19 disease by using certain protein mapping techniques and a deep learning. The proposed method consists of 5 stages. In the first stage, drug-target interactions were obtained from the DrugBank database. In the second stage, mapping of drug compounds and target proteins was made. While PubChem fingerprinting method was used for the mapping of drug compounds, target proteins were mapped with 6 different methods; Meiler parameters, Atchley factors, PAM250, BLOSUM62, Miyazawa energies and Micheletti potentials. In the third stage, the mapped drug compounds and the mapped target proteins were combined and a one-dimensional feature space was obtained. In the fourth stage, the one-dimensional feature that was generated before was classified with the LSTM (Long-Short Term Memory) deep learning model and the prediction was performed. In the last stage, the performance of the protein mapping methods was determined and compared with accuracy, precision, recall, f1-score, and ROC (Receiver Operating Characteristic) evaluation matrices. When the application results were examined, it was seen that all protein mapping techniques performed above 85%. The best accuracy and ROC scores were obtained from Atchley factors and Meiler parameters. With Atchley factors, an average of 92% accuracy and 98% ROC were obtained. With the Meiler parameters, the ROC value did not change, but the accuracy value was measured as 91%. Afterwards, it was observed that Micheletti potentials and Miyazawa energies showed the second best performance. On average, 90 and 91% accuracy values were obtained, respectively. ROC values were calculated to be close to each other and 98% ROC value was obtained for Micheletti potentials, while this ratio decreased to 96% with Miyazawa energies. BLOSUM62 and PAM250 protein mapping methods were more ineffective than other methods. While BLOSM62 showed an average accuracy of 87%, PAM250 predicted drug-target interactions an average of 91% accuracy. While the ROC value of the BLOSUM62 method was 89%, this rate increased in PAM250 and a ROC value of 92% was obtained. Contributions obtained by the end of the study can be expressed as follows; with this study for the first time, drug-target interactions of COVID-19 were predicted by protein mapping techniques. In addition, the most effective protein mapping method among protein mapping techniques was determined. It was demonstrated that the selected protein mapping techniques are important in determining drug-target interactions. Additionally, it has been observed that computational-based methods can be at least as effective as experimental approaches. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG. | en_US |
