Unsupervised Prediction Method for Drug-Target Interactions Based on Structural Similarity

Abstract

Predicting drug-target interactions are important in drug discovery and drug repositioning. Discovering drug-target interactions by computational method still has great potential and advantages in many aspects, such as focusing on interested drug or target proteins. This paper proposes an unsupervised clustering model (OpBGM) based on structural similarity, which combines OPTICS and BGMM algorithms. First, the required PDB files are obtained. Then, the interaction pair is defined and extracted from each PDB file. The interactions are encoded and dimensionally reduced using PCA algorithm. OPTICS is used to detect and remove noise, and BGMM is used to extract significant interaction pairs. Potential binding sites are discovered through interaction pairs and drug similarities discovered. In addition, a target protein is randomly selected to dock with each drug in one cluster. The number of clusters with the average affinity less than −6 kcal/mol accounts for 82.73% of the total number of clusters, which shows the feasibility of proposed prediction method.