An Enhanced Computational Approach Using Multi-kernel Positive Unlabeled Learning for Predicting Drug-target Interactions

Abstract

Background:: In recent years, analyzing complex biological networks to predict future links in such networks has attracted the attention of many medical and computer science researchers. The discovery of new drugs is one of the application cases for predicting future connections in biological networks. The operation of drug-target interactions prediction (DTIP) can be considered a fundamental step in identifying potential interactions between drug and target to identify new drugs. Objective:: The previous studies reveal that predictions are made based on known interactions using computational methods to solve the cost problem and avoid blind study of all interactions. But, there seem to be challenges such as the lack of confirmed negative samples and the low accuracy in some computational methods. Thus, we have proposed an efficient and hybrid approach called MKPUL-BLM to manage some of the aforementioned challenges for predicting drug-target interactions. Methods:: The MKPUL-BLM combins multi-kernel and positive unlabeled learning (PUL) approaches. Our method uses more information to increase accuracy, in addition to minimizing small similarities using network information. Also, potential negative samples are produced using a PUL approach because of lacking negative laboratory samples. Finally, labels are expanded via a semi-supervised. Results:: Our method improved to 0.98 and 0.94 in the old interactions set for the ROCAUC and AUPR criteria, respectively. Also, this method enhanced ROCAUC and AUPR criteria by 0.89 and 0.77 for the new interactions set. Conclusion:: The MKPUL-BLM can be considered an efficient alternative to achieve more reliable predictions in the field of DTIP.