Machine learning based approach for large-scale drug-target binding prediction

Abstract

Identification of targets to drug molecules is very important in understanding how drugs work in the human body. In particular, recent developments in phenotypic screening have led to increasing attempts to select actual targets from large protein targets. In addition, multiple target prediction is essential for drug repositioning and prediction of side effects of drugs in advance. Conventional methods for identifying targets require a lot of time and money, so virtual screening using machine learning techniques are gaining popularity. Structure-activity relationships (SAR), a well-known method of identifying targets, has the advantages of low computational costs and high availability, while risking biased results due to data dependencies. In this paper, we first introduce a multi-target prediction algorithm based on a random forest model. The model features performance optimization through various data preprocessing methods to overcome data bias. In addition, this paper introduces “Multiple Partial Multi-task learning (MPMT)” which improves the prediction performance through analysis of the deep learning methods used in the field of drug-target binding prediction.