Dextrosinistral reading of SMILES notation: Investigation into origin of non-sense code from string manipulations

Abstract

The SMILES notation provides a digital way to represent any chemical structure in the form of a string of ASCII characters, therefore, a preferred data medium for machine learning models. As Chomsky type-2 language, SMILES notation is supported with context-free grammar, raising errors for invalid string arrangements. Numerous efforts have been made to recover chemical structures in invalid SMILES strings. Exploring the flexibility of SMILES notations of real molecules would give critical information related to SMILES string reorganizations and sources of errors. Present study examined the potential for reading SMILES notation from right-to-left, known as dextrosinistral reading, and evaluated the effect of new character combinations on the representative chemical structures. The study developed a set of string operations to reverse the order of characters in the SMILES string while maintaining the context-free grammar of SMILES notation. These operations were tested on SMILES notation of over two hundred natural products, resulting in diverse changes at the chemical structure level, including reverting to the original structure, reconfiguring into an isomeric structure, or generating compounds having valency errors. The DFS-tree profiled the changes in chemical structures from reorganizations of SMILES strings and identified the source of atoms with valence errors. Molecular Mechanics (mm2) calculations showed that a group of newly generated chemical structures has total energy in a range of transition state molecular complexes. While the analyses of machine learning models showed the need for cheminformatics tools, such as RDKit and OpenBabel libraries, to develop modules that can fingerprint the reorganized SMILES strings containing atoms of explicit valences. The outcome of the present study highlighted the diversity and flexibility of SMILES notation, and may provide a new source of data required for developing the cheminformatics functionalities necessary to advance machine learning-based chemical discovery.