Representation of Molecules by Sequences of Instructions

Abstract

The processing of chemical information by computational intelligence methods faces the challenge of the structural complexity of molecular graphs. These graphs are not amenable to being represented in a suitable way for such methods. The most popular representation is the SMILES notation standard. However, it comes with some limitations, such as the abundance of nonvalid strings and the fact that similar strings often represent very different molecules. In this work, a completely different approach to chemical nomenclature is presented. A reduced instruction set is defined, and the language of all strings that are sequences of such instructions is considered. Instructions provide the means to incrementally add atoms and modify the connectivity of the chemical bonds of atoms to be inserted. Instructions are carefully crafted to guarantee that all strings of this language are valid, i.e., each string represents a molecule. Moreover, slight changes in a string usually correspond to small modifications in the represented molecule. Therefore, this approach is appropriate for use in state-of-the-art computational intelligence systems for chemical information processing, including deep learning models.