Study Of SNPs In Human Cytochrome P450 And Protein-chemical Interactions

Personalized medicine is the tread of modern healthcare. In this thesis, polymorphism in cytochrome P450 was studied for drug target researches. And protein-chemical interactions were predicted for studies of drug-ligand pairs.The cytochrome P450(CYP450) is a super family of enzymes which responsible for multiple metabolic processes. Most xenobiotics and endogenous compounds, including more than 90% therapeutic drugs, are metabolized by the cytochrome P450 enzyme system. Polymorphisms in genes of CYP450 s can change the enzyme activity, like complete deficiency and ultrafast metabolism. In drug metabolism, the differences lead to severe toxicity or therapeutic failure. Furthermore, various gene polymorphisms have been associated with various diseases. SNPs are the most common forms of polymorphisms in CYP450 s.In silico, some studies were carried out to predict SNPs by developing algorithms based on base pair composition. As the limited information they extracted, the prediction accuracies are around 50%. We have therefore proposed three algorithms to predict SNPs. The first proposed model is SCYPPred, which predicts human cytochrome P450 SNPs(Single Nucleotide Polymorphisms) based on SVM and flanking sequence method. SCYPPred can rapidly yield the desired results with 66.7% accuracy by using the amino acid sequences information alone. In the second prediction model, we put forward an effective representation of SNPs by modelling candidate residues and flanking DNA sequence, including information on the sequence composition, existent SNPs and “CpG dinucleotides”. Using combined features, we further trained SVM classifier and gained an accuracy of 75.56%. The third model is used a large range of features, from the information of sequence composition to the attributes of target site and evolutionary information were collected. And a new method which based on fuzzy set theory to balance the datasets was applied. After SVM training and test, we generated a model with 92.5% in accuracy, which showed remarkable improvement in prediction performance.Drug is an effective solution for diseases treatment. And identification of protein-chemical interactions is a crucial stage of drug discovery. Computational methods are more capable to handle considerable data and efficient to detecting uncovered interactions between compounds and proteins.Here, we developed an algorithm to predict protein-chemical interactions with both high accuracy and large coverage by applying Bayesian additive regression trees(BART) on a novel proposed uniform space termed bow-pharmacological space. Bow-pharmacological space is formed by three parts: protein space, chemical space and chemical-protein interaction space. Furthermore, in our prediction, we extended data to large scale which includes all proteins in human, and chemicals in STITCH database. In the results, our prediction model demonstrates excellent performance which is 98.56% in accuracy. We also carried out a case study to search the potential ligands for Kinesin-like protein KIF11. From our prediction, G7 X is a new interacted compound, which is also verified by a latest experiment and our docking analysis.