Research On Modeling Methods Of Quantitative Structure-activity Relationship For Environmental Chemical Toxicity

The large number of compounds present in the air, soil and water, and manyother environmental elements, the determination of their qualitative and quantitativetoxicities to human, animal and plant is an urgent problem needed to solve. Thecurrent testing means of the toxicities of environmental compounds is animalexperiments, where usually the cheap and fast in vitro experiments are used forprescreening, and the expensive and time-consuming in vivo experiments for finaltest. The biggest problem faced by animal experiments is ethical issue, and with theimprovement on human civilization and the deeper understanding upon therelationship between human beings and their co-residing Earth’s flora and fauna, theethical issue would be a biggest problem facing animal experiments; secondly,animal experiments especially in vivo experiments, owing to their high time cost andhigh monetary costs, also limit the quantity of compound for testing in practice. Toalleviate such bottleneck of animal experiments, QSAR technology appeared andgradually grew up, it involving mathematics and statistics, quantum mechanics,biology, and computer science, is the causal relationship model between molecularstructure and toxicity of compound. QSAR is based on mathematical and statisticaltheory to model, uses computer science as a tool to realize the mathematical andstatistical theory, uses quantum mechanics as a tool to acquire the molecular structureof compound, uses biology as a tool to acquire the toxicity data and theunderstanding of toxicity mechanism of compound, by QSAR model the toxicityvalue of compound can be obtained directly from molecular structure of compoundwithout animal experiments. It’s this possibility that QSAR technology would canreplace animal experiments to be a testing means to compound toxicity, that has ledto a great impact upon the current testing technology for compound toxicity by QSAR, and it would be forecasted that, QSAR would further exert profound impactupon the future development of current testing technologies.This dissertation targets at the test for toxicity of environmental compounds,uses QSAR technology as the testing means, explores the possibility of QSAR to bea testing means to compound toxicity instead of animal experiments, a total of threeQSAR models are built, respectively carcinogenic classification model, estrogenreceptor binding capability classification model and cerebral blood barrierpermeability classification model, the performances of three models are evaluated bymeasurements of animal experiments.Firstly, using the molecular structures and the long-term rodent animalcarcinogenicity bioassays values of1153environmental compounds provided by U.S.Environmental Protection Agency, build the carcinogenicity classification model ofenvironmental compounds. Based on the assumptions that molecular descriptorsfollow normal distribution and compound toxicity classification values followbinomial distribution, Rogers distribution function is constructed which the whole ofmolecular descriptors and toxicity classification values of1153compounds follow;use Laplace premise to transform negative logarithm likelihood function so that thebalance between the contradicting sparsity and fitting of Rogers distribution is gained;use cross-check to select9molecular descriptors from the sequence ordered byweights of729molecular descriptors, which serve as the structure data ofcarcinogenicity classification mode of compound to be built; by maximizing thedistance between negative and positive carcinogenic compound which is theoptimizing condition, select797compounds from all1153compounds as the supportvectors, select Gaussian kernel function to measure the relativity between pairwisecompounds, use support vector machine to construct super plane which classifies1153compounds into two classes of negative and positive; use the long-term rodentcarcinogenicity bioassays values of1153carcinogenic compounds to evaluate theperformance of classification model built, the classification accuracy rate of1153compounds in carcinogenicity by model is66.86%.Secondly, using the molecular structures and the rat uterine cytosol estrogenreceptor competitive binding experimental values of278environmental compoundsprovided by U.S. Environmental Protection Agency, build the estrogen receptorbinding capacity classification model of environmental compounds. Use entropy to construct the symmetry impermanence of compound, use symmetrical impermanenceto measure simultaneously the causality between molecular descriptor and estrogenreceptor binding capacity of compound as well as the redundancy between pairwisemolecular descriptors; design an algorithm to select8high causal and low redundantmolecular descriptors from729molecular descriptors of278compounds, whichserve as the structure data of estrogen receptor binding capacity classification modelof compound to be built; construct a8-dimensional Cartesian feature space, useEuclidean distance to measure the pairwise similarity of278compounds, by knearest neighbor use4most similar compounds in structure to determine the negativeor positive of estrogen receptor binding ability of compound; use the rat uterinecytosol estrogen receptor competitive binding experiment values of278compoundsto evaluate the performance of estrogen receptor binding capacity classificationmodel of compound built, the classification accuracy rate of278compounds inestrogen receptor binding ability by model is96.76%.Finally, using the molecular structures and the cerebral blood barrierpermeability in vivo measurement values of80environmental compounds providedby QSAR World, build cerebral blood barrier permeability classification model ofenvironmental compounds. Construct the complete graph of all80compounds, usedot product to calculate adjacency matrix, scale matrix and Laplace matrix, usesingular value decomposition to obtain the eigenvalue and eigenvectors of Laplacematrix, use the spectrum of full graph to measure the goodness of moleculardescriptor; use cross-check to select9molecular descriptors from the sequenceordered by goodnesses of729molecular descriptors, which serve as the structuredata of cerebral blood barrier permeability classification model of compound to bebuilt; construct Bayesian classifier to serve as the cerebral blood barrier permeabilityclassification model of compound, use the naivety assumption to transform jointprobability into independent probability, use frequency to calculate the negative andpositive probability of cerebral blood barrier permeability of compound, use normaldistribution to construct the probability distribution of molecular descriptors, usemaximum likelihood estimate to obtain the mean and variance values of normaldistribution; use the cerebral blood barrier permeability in vivo measurements of10compounds to evaluate the performance of cerebral blood barrier permeability classification model of compound built, the classification accuracy rate in cerebralblood barrier permeability by model is90.00%.