Year: 2014
Type: dissertation
Abstract: Small molecules that have the ability to alter the biological response of a cell or disease state are of significant interest as drug candidates.One approach to the design of such molecules using kernel methods is through a map of molecules to a feature space induced by a kernel, where the predictions of the molecular properties are made in a linear manner.The forward mapping helps guide the synthesis of new molecules.Another more direct approach is to solve the so-called pre-image problem, i.e., to reconstruct a corresponding molecule (aka pre-image) from its feature space representation.The inverse mapping is of central importance for the design of new molecules with desired properties.In this thesis, we address two problems in drug design: (I) the forward problem of predicting molecular properties, where we propose a new graph kernel that induces a feature space amenable to the prediction of molecular properties, and (II) the inverse problem of designing new molecules that possess properties required for drug candidates, where we develop a population-based Monte Carlo method to solve the pre-image problem for the molecules.Our respective contributions to these problems are summarized as:The measurement of molecular similarity is an essential part of predicting molecular properties.Graph kernels provide good similarity measures between molecules.A conventional graph kernel is based on counting common subgraphs of a specific type in molecular graphs.This approach suffers from two primary limitations: (i) only exact subgraph matching is Many thanks go to my thesis supervisor, Ryo Yoshida, and sub-supervisor, Tomoyuki Higuchi, for introducing me to the world of machine learning.Ryo has helped me in innumerable ways.
Authors: 博史 山下, ヒロシ ヤマシタ
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W2315509919