Preimage Reconstruction of Molecules with a Graph Kernel

• Activity Number: 187
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Ryo Yoshida*+ and Hiroshi Yamashita and Yukito Iba and Tomoyuki Higuchi
• Companies: The Institute of Statistical Mathematics and Graduate University for Advanced Studies and Institute of Statistical Mathematics and Institute of Statistical Mathematics
• Keywords: Graph kernel ; cheminformatics ; MCMC ; chemical compound ; preimage problem

Abstract

With a newly-derived graph kernel, some problems on chemical informatics are addressed. For a pair of compound graphs, most existing kernels are designed to count commonly-appeared molecular fragments where only the fragments sharing exactly the same atomic compositions and covalent bonds are taken into account. Accordingly, a pair of fragments with one or few structural variations is never reflected. This limitation is addressed. Each vertex is annotated by an attribute of the atom reflecting topological and physicochemical environments. The new kernel is designed to measure the closeness of the attributes for all paired atoms with their possible arrangements. With this kernel, we tackle the two problems: (i) Developments of classifiers to predict several types of biochemical or physiologicalproperties; (ii) Preimage reconstruction of molecules that explores the inverse mapping of a graph in the RKHS to the original chemical space. The main contribution of this study lies in the second problem. To solve the preimage problem, we present a new MCMC method where randomly-chosen subcomponents of a compound are replaced successively to molecular fragments collected in a database.