Research On Ecological Intrinsic Properties Prediction For New Chemicals

Inherent properties of chemical compounds are of great importance to environmentinfluence judgment and ecological risk assessment. Traditionally, the data of ecologicalinherent properties are mainly obtained by experiment, which is costly, with large workloadand time-consuming. In order to accurately evaluate the ecological harm of compounds intime and reduce the number of experiments, in this thesis, prediction methods of basicecological inherent properties including octanol-water partition coefficient, degradability,bioconcentration factor and fish toxicity based on quantitative structure-activity relationshipare studied.Firstly, the fragment structures of compounds are recognized. After that, predictionmodels of the mentioned above four properties are constructed based on principal componentsanalysis (PCA)-multivariate linear regression (MLR), principal components analysis(PCA)-backpropagation network (BP) and principal components analysis (PCA)-supportvector machine (SVM). And finally, an intelligent properties prediction system is developed.The main contents and conclusion are listed as follows:(1) Construction of fragment structures set and recognition of these structures. Under theguidance of experts from Nanjing Institute of Environmental Science, the main effect factorsof above four properties are analyzed and selected. And with these effect structures,corresponding fragment structures sets are established for modeling. And then, based on therules of SMILES (simplified molecular input line entry specification) and with thecompounds information from Nanjing Institute of Environmental Science and website as datasources, fragment recognition algorithm is designed which can accurately recognize thefragment structures of the set. Besides, the recognition result is verified by referring to manualidentification result.(2) Construction of four properties QSAR models based on PCA-MLR, PCA-BPandPCA-SVM methods. In order to eliminate the redundancy between the various fragmentstructures, before construct the model, principal component analysis method is used to filterthese structures, fragment structures which impact largely on the characteristics are selectedas the model input parameters. With each property data as output parameter, respectively usethe above three methods and select the appropriate model parameters to establish QSARmodels of the above four properties, and finally used cross validation and external validation methods to verify the established models. The validation results indicate that these models canpredict the corresponding property data scientifically and accurately.(3) Design and development of a prediction system for ecological inherent propertiesbased on JAVA and MATALB mixed programming technique. A prediction system is realizedby integrated the recognition result of fragment structures and the constructed models of fourproperties. It can not only predict these four properties accurately, but also can export theprediction results with Word, Txt and Excel. Besides, chemical two-dimensional structureformation is realized by using CDK (Chemistry Development Kit) component, which canshow the chemical structure of prediction compounds visually. A series of tests prove that thesystem is of good stability and applicability.