Research On Data Mining And Prediction Of Drug Interactions

Drug discovery is a long process which can spend a lot of time and cost a lot of money,for example,a successfully novel marketed drug always took about decades and costed nearly billions dollar.Meanwhile,more and more data in drug development has been explosively increased.It has been a daunting issue to address the massive data by conventional experiments.Nowadays,a growing number of in-silico methods have been developed to aid experimental methods to solve drug development problems.This dissertation mainly focused on drug interactions in drug discovery.The potential relationships of drugs and targets were excavated and analyzed,so the corresponding predictive models were constructed.Among them,one was developed to predict hot spots based on traditional machine learning algorithm,and the others were based on deep learning methods,involving the prediction of drug-target interactions and drug active screening by QSAR techniquey.The first step in drug development is identifying target positions.Hot spots are key residues on the core areas of protein-protein binding interfaces,obviously influencing the structure and function of protein.They are usually regarded as target positions and applied in the field of drug design.Recently,in-silicon computational methods have been widely used for hot spot prediction through sequence or structure characterization,since experimental methods are costly and time-consuming.As the structural information of most proteins is unavailable,identifying hot spot simply from amino acid sequences presents more useful and meaningful for real-life applications.This dissertation proposed a new sequence-based model that combines physicochemical features with the relative accessible surface area of amino acid sequences for hot spot prediction.The ensemble of several individual classifiers produced final prediction.After the target positions of drugs were identified,discriminating drug-target interactions(DTIs)is also an important step in drug development.The DTIs are anything within a living organism to bind with some other entities(like an endogenous ligand or a drug),resulting in change in their behaviors or functions and thus being able to treat diseases.Here,a novel deep-learning-based model was proposed for discriminating DTIs.The predictive performance was further improved owing to extracting deeper and subtler features by deep-learning-based model.Moreover,a relative reasonable method for generating negative instances was also proposed in this dissertation,where negative instances were evaluated by the distance between all positive instances and possible negative instances.The larger distance between one negative instance and all positive instances means that the negative sample is more credible.The experimental results indicated that the predictive model yielded successful performance on two different datasets and thus it can be employed to discriminate the interactions between drugs and targets.The last step in drug development is screening lead compounds.The research on quantitative structure-activity relationships(QSAR)is actually a ligand-based virtual screening method,which provides an effective approach to accurately determine new hits and promising lead compounds during the drug discovery process.In the past decades,various research works have gained highlighted performance with the development of machine-learning algorithm.The rise of deep learning technique,along with massive accessible chemical databases,has resulted in improving the QSAR predictive performance.Therefore,a novel end-to-end deep-learning-based model was designed to implement QSAR virtual screening and two distinct training scenarios were utilized to assess the predictive model.The encoder-decoder model was mainly used to generate fixed-size latent features to represent chemical molecules;while convolutional neural network(CNN)framework used these features as input vectors to train a robust and stable model.Both two scenarios have verified the validity of predictive model.In summary,the purpose of this dissertation was to build robust predictor to address the issues about drug interactions.Drug interaction data was mined and analyzed through machine learning or deep learning algorithms,discovered their potential relationships and then constructed stable and robust predictive models.Compared with other state-of-the-art models,the proposed models produced satisfactory performance.Thus,the proposed predictive models can be useful to aid relative experiments and provide a highlighted sights into experimental results.All in all,the work can be an effective way to shorten drug development time and reduce the cost of research.