Representation Of Chemical Compound And Its Application To Chemoinformatics

People are not interested in beautiful chemical structure, but rather in its properties. Hence, the firstfundamental task in chemistry is to make inferences about which structure might have the desiredproperties. The first work of model construction of structure-property or structure-activity relationship isthe representation of chemical compounds, that is, the development of molecular descriptors. This thesisconsists of2D molecular shape description, molecular code based on chemical bonds and their applicationsin chemoinforatics. The details were as follows.1. The studies on2D molecular shape and its chirality description.The chirality description of compounds is an essential subject, so far several reasonable and effectivealgorithms derived from a diversity of strategies have been approached to represent chirality of molecularstructures in3D. However, the concept of chirality has been extended to n-dimensional space, including2D. If a molecule embedded in a plane, such as benzenoids, cannot be translated to coincidence with itsmirror image in the plane, the molecule is chiral in2D, although it is not chiral in the3D space.In order to represent atomic asymmetry, we suggested the methods of weighed atomic sums on thebasis of binary code of molecular periphery for the descriptions of atomic asymmetric environments. Theweighed atomic sum of an atom is a measure of the asymmetry of the atomic environment as obtained bycirculating around the molecule going in opposing directions, and so the weighed atomic sum of an atom ina molecule and that of corresponding atom in the opposite enantiomer are of the same magnitude but of theopposite sign. Weighed atomic sums were available not only chiral molecule, but also achiral molecule.Compared with atomic sums approached by Randi, weighed atomic sums are more distinctive and better than atomic sum to represent atomic asymmetry. For example, the asymmetric atoms which were wronglyassigned as symmetric atoms by atomic sums were correctly described by weighed atomic sums.In addition, the chirality degrees of planar bezenoids were suggested on the basis of weighed atomicsums which are the descriptions of atomic asymmetric environments related to the distance between atoms.The chirality of compounds was quantitatively measured, rather than labeled as chiral or achiral. Theenantiomers could be distinguished from each other and the values of their chirality measurements takeopposite number. The chirality degrees could be a multi-dimensional vector or a single value to representchirality of compounds. The three dimensional structures of n-alkane rotamers could be transformed into2D benzenoids, and then the chirality degree also could be used to describe the3D conformers.2. Studies on MOLMAP descriptors and its applications.The molecular mapping of atom-level property （MOLMAP) descriptor was generated on the basis ofchemical bond descriptors of a molecule by Kohonen self-organizing Map （Kohonen SOM) with aspecific algorithm. The bond descriptors were composed of physiochemical properties of the chemicalbond, such as the difference of the charges between the two atoms and topological properties, such as thenumber of hetero-atoms connected with the two atoms. In the present study, the MOLMAP descriptorswere implemented to predict the mutagenicity of4075organic substances （2305mutagens and1770nonmutagens in Ames test). The random forests were used to construct mathematical models by threekinds of descriptors:(1) MOLMAP descriptors of different size;(2) global molecular descriptors;(3) thecombination of MOLMAP descriptors and global molecular descriptors. The correct predictionpercentages of out of bag (OOB) cross-validation of the whole data set were reached to85.4%. In order totest the stability of the prediction model, the built model was used to predict the test set which wascomposed of472compounds collected from another database. The percentage of the correct prediction of the test set was86.7%. To compare with the results of the previous work, both of the prediction resultswere improved.