Research On The Compound Activity & Selectivity Based On RBM

Searching for new drug from nature is a time-consuming process. But with the development of computer-aided drug design and cheminformatics, both Structure-Activity-Relationship (SAR) and Structure-Selectivity-Relationship (SSR) earn widespread respect; this greatly reduces the cost of drug development. However, there is still a lack of efficient and accurate computational methods to predict compound activity properties; and the selectivity test can include binding assays or clinical trials which defers selectivity assessment to the later stages, so if it fails, then significant investments in time and resources get wasted. These are the bottleneck of drug discovery and also a challenge for the speed and accuracy of new drug development.Compound activity and selectivity prediction based machine learning is always an important topic of drug discovery, which have been done a lot of research. But these SAR and SSR researches mainly focus on the statistical analysis, and in this paper we develop machine learning approaches such as collaborative filtering based on Restricted Boltzmann Machine (RBM) and Deep Learning (DL) to predict and analyze the activity and selectivity of compound. The main research of this paper is organized as follows:1. We analyze the analogy between the data used for recommender systems and chemical informatics. Then we build and get a target-compound matrix which is sparse.2. We expand RBM to the collaborative filtering for chemistry problem, and predict compound activity.3. Improve the Deep Belief Networks (DBN) algorithm with Multitask Learning (ML), which learn the distributed representation of compound using RBM unsupervised learning process and promote the accuracy of selectivity prediction using ML approach.In this paper, we will test the accuracy and efficiency of collaborative filtering for chemistry problem based on the target-compound matrix dataset. And test the accuracy and efficiency of DBN with ML based on the SSR dataset collected by the University of Minnesota. Our results will greatly reduce the cost and shorten the cycle of drug discovery, moreover, it will help the chemical scientists to discover new drug from the massive data.