A new approach to atom-to-atom mapping using the naive Bayesian classifier

• Published in: Uchenye zapiski Kazanskogo universiteta. Serii͡a︡ Estestvennye nauki Vol. 160; no. 2; pp. 200 - 213
• Main Authors: A.I. Khayrullina, T.I. Madzhidov, R.I. Nugmanov, V.A. Afonina, I.I. Baskin, A.A. Varnek
• Format: Journal Article
• Language: English; Russian
• Published: Kazan Federal University 01-06-2018
• Subjects: atom-to-atom mapping; chemical reactions; chemoinformatics; classification; machine learning
• ISSN: 2542-064X; 2500-218X

The key step in the computer analysis of chemical reactions is the determination of the correspondence between the atoms of reagents and products. This procedure is called atom-to-atom mapping (AAM). The presence of AAM is a key factor for establishing the mechanism and type of reaction, searching for similarities and substructures, modeling, checking the quality of data. A new approach has been proposed to the search for optimal atomic-atom mapping in chemical reactions based on the use of machine learning methods. The learning task is formulated as a classification: for each pair of the reagent-product atom, it is necessary to establish their assignment to the correct/incorrect mapping. We have used a simple naive Bayesian classifier. The approach described in this paper is the first example of a self-learning algorithm for AAM.