Research On Prediction Of Drug-likeness And Activity Under Big Data Environment

For improving efficiency of drug design, better quality lead compounds need to be screened out from compound database. Therefore, the concept of drug- likeness, was proposed by pharmaceutical chemists, and developed predictive rules of drug-likeness based on characters and properties of molecular structures. In addition, research on relationship between compound structure and activity is also an important approach of drug design, and it’s of equal importance to the drug-likeness during process of new drug research and development. Most traditional studies of compound activities determining drug activities by biological assay and living animal experiment, and had to spe nd lots of sources and time. Modern researches on compound activity use mathematical methods to build quantitative structure-activity relationship model to predict activities of unknown compounds. With the development in computer data mining technology, machine learning has been an active research direction in computer science at present, thus scientists could apply it to improve predictive efficient of compound activity. However, most of existing studies used shallow machine learning methods, so that expressing abilities to complex functions are restricted in the case of finite samples and computing units, and can’t learn more useful features. Also numbers of samples in these researches are small and accuracies are low, and practicability would be less under big data environment.In this paper, deep learning method is used to build predictive model of drug- likeness and activity based on mass compounds data. Paper includes the following two aspects:(1)Predictive model of drug- likeness for mass compounds data. In this work, a distributed computing model for fast retrieval of massive compounds is built according to special data structure of compounds. Then screening compounds that may drug- like quickly and efficiently according to rules of drug- likeness.(2) Predictive model of activity for mass compounds data. Shallow and deep machine learning methods are discussed in this work, and analyses pros and cons of supervised and unsupervised learning. Then uses molecular descriptors as features, to build predictive model of activity for mass compounds data by different machine learning methods respectively.The experimental results show that the proposed deep learning model can be applied to predict drug- likeness and activity of mass compounds, and can screen drug- like compounds rapidly and predict their drug activities. The model is stably scalable and efficient, and the accuracy has also been reflected.