PULBLM: A Computational Positive-Unlabeled Learning Method for Drug-Target Interactions Prediction

Drug-target interactions prediction is an essential prerequisite for drug discovery and human medicine. Computational drug-target interactions prediction methods are low-cost alternatives to laboratory activities. One of the most popular computational methods for drug-target interactions prediction is bipartite local models (BLM). This method uses local models to predict potential interactions based on drug domain similarities, target domain similarities, and known drug-target interactions. One of the challenges in this field is the absence of confirmed negative laboratory samples. Using positive-unlabeled learning (PUL) approaches to extract reliable negatives can help improve performance. In this article, we have expanded BLM by introducing a PUL method. We have used real-world online datasets to test our method. The results shows that our method has better results than BLM and SELF-BLM. Our method has 0.82 and 0.59 results in AUC and AUPR, respectively.