Machine-learning (ML) technologies are expanding the boundaries of molecular de novo design. This cooperative effort allows increasing the searchable chemical space, whilst decreasing time and expenses related to chemical synthesis. This work seeks to use ML technologies to design synthetically feasible novel molecules, with suitable pharmacological and toxicological profiles and potential activity towards targets relevant in diabetes mellitus. This criteria will serve as an objective function to maximize in a multi-target optimization problem.<br/><br/>Character-by-character generation of SMILES by deep recurrent networks requires careful hyper-parameter initialization since catastrophic forgetting might occur for complex optimization targets. To overcome this problem we introduce a novel molecule optimization framework, which characterizes SMILES representation as a tree structure. Since the SMILES' sintaxis can be specified as a context-free grammar, it provides a definition of what can or can not be permitted in the generation of the SMILES strings. Using this specification, a deep neural network to learn how to properly generate SMILES is no longer needed. This strategy allowed us to only focus on maximizing the objective function, using the syntactic rules the grammar provides.<br/><br/>This tree structure will serve as the main component in a genetic algorithm optimization: we select promising molecules, then we recombine and mutate them to produce better fitted individuals. This tree can be parsed using the syntactic specification of the SMILES, which allows for a rich definition of genetic operators, while minimizing the chances of getting invalid structures.

Molecular de novo design using context-free grammars

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.