Development of drug-target interaction prediction model using deep learning and various molecular fingerprints

Abstract

In drug discovery, the identification of drug target interaction (DTI) plays a key role. However, it takes a lot of cost and time for DTI identification through experiments. Computational DTI prediction can solve this problem because it can help to narrow down the drug candidates and it reduces cost and time for drug discovery. There are various approaches to computational DTI prediction and one of the popular approaches is a feature-based approach. In the feature-based approaches, the input features for drugs and target proteins are very important and there are various types of features. In this study, I more focused on drug features and one of the popular drug features is a molecular fingerprint. To find the best molecular fingerprint for DTI predictions, I compared DTI prediction performances using deep learning and various molecular fingerprints and the model which used Mol2vec had the best performances in F1-score and AUPR. Additionally, to improve prediction performances, I used multiple molecular fingerprints for DTI prediction. The model which used Mol2vec and Seq2seq fingerprint simultaneously had the best performances in F1-score and the model which used Mol2vec and Neural fingerprint simultaneously had the best performances in AUPR. Also, I compared proposed models with baseline models and proposed models had better performances in DTI predictions. Furthermore, I predicted drug properties using various molecular fingerprints to find reasons why the model which used the specific molecular fingerprints had higher DTI prediction performances. According to drug property prediction results, the model which can describe binding information well had higher DTI prediction performances and the model which can describe general drug properties well had higher DTI prediction performances. Finally, I found features of proposed models which had higher DTI prediction performances. The model which used continuous molecular fingerprints that were extracted from pre-trained models had higher DTI prediction performances. Also, Low dimensional molecular fingerprints had better performances in DTI prediction than high dimensional molecular fingerprints.