Research On Predicting The Interaction Between Receptors And Durgs Based On2D Molecular Fingerprint And Imbalanced Dataset

In the process of developing new drugs, traditional methods have obviousshortcomings such as long period, great expense, strong blindness, unknown mechanismof drug. With the completion of human genome project and the rapid development ofbioinformatic technologies, it promotes the development of the computer-aided drugdesign.Identification of drug-target interactions is a key step in the drug discovery processof the computer-aided drug design. However, biological experiment approaches wastetime and energy and now computational approaches study drug development mostlybased on structure-based drug design, which relies on the knowledge of thethree-dimensional structure of the target protein. However, the number of the unknownstructure protein is far more than that of the known structure protein. For the unknownstructure proteins, it is not versatile and accurate to create a homology model. Therefore,as a complement, it would be beneficial to study the drug-target interactions based onsequence-derived features via computational methods.Here we propose a new predictor for identifying enzyme-drug interactions based onsequence-derived features and molecular fingerprints, combining machine learning andpattern recognition theory. In the process of drug design, some proteins such as enzymes,ion channels, G protein coupled receptors (GPCRs) and nuclear receptors represent themajority of the current drug targets, which are the most successful applied objects inhuman body. Therefore, this paper analyzes and compares the interaction between thesetarget proteins with drug carefully, providing the efficient and reliable supportingresearch for biopharmacy. The main work and the possible creative achievements in thisthesis are shown as followed:(1) Based on the protein theory and corresponding knowledge, for constructingdigital model problem about amino acid sequence, we fused different kindsof amino acid sequence information such as pseudo amino acidcomposition, PSSM matrix, physicochemical property, dipeptidecomposition and grey dynamic factor, etc. This model is not only simple,but contains a wealth of information about physical-chemical property andgenetic evolution.(2) For digital description method research based on molecular fingerprintabout small molecular structure of drug, we represented drug structure witha discrete numerical sequence. This method not only includes molecularstructure information, but has the advantage of quick calculation. The results shows it have a good performance.(3) Designing four classifiers for target protein, whose results perform betterthan others, and carefully researching the results about using imbalanceddataset and not using imbalanced dataset.(4) Constructing the user-friendly online predictor web-server, detailedinstructions and ease of use, this is particularly useful for mostexperimental scientists to obtain their desired data in a timely manner.