Modélisation de la stéréochimie : une application à la chémoinformatique
Pierre-Anthony Grenier · HAL (Le Centre pour la Communication Scientifique Directe) · 2015
In the framework of prediction of molecule’s properties, graph kernels allow to combine anatural encoding of a molecule by a graph with classical statistical tools. Unfortunately somemolecules encoded by a same graph and differing only by the three dimensional orientations oftheir atoms in space have different properties. Such molecules are called stereoisomers. Thiswork aims to take into account stereoisomerism into graph kernels method. In this documentwe first present the main methods of prediction of molecule’s properties, and we focus onmethods based on graph kernels. Based on this state of the art, we present stereoisomersand a state of the art of methods which take into account those molecules. Then we proposeto encode stereoisomers by ordered graphs. We define minimal stereo subgraphs, which aresubgraphs that locally characterizes the stereochemistry. Those subgraphs are used to definea kernel that take into account the stereochemistry. Finally we propose three extensionsfor this kernel. Those extensions allow to consider the neighbourhoods of minimal stereosubgraphs and to compare different minimal stereo subgraphs.