Industrial Case Study: Identification of Important Substructures and Exploration of Monomers for the Rapid Design of Novel Network Polymers with Distributed Representation

Abstract Designing polymers experimentally is a time-consuming task. Quantitative structure-property relationship analysis can help speed the development of new polymers. The authors hypothesized the ideal mixture model, with which polymers are represented by composition-weighted descriptors of monomers. In this study, we pursued a new polymer that had the desired properties from an industrial dataset. We first constructed a partial least squares (PLS) model and random forest with five descriptor sets. The PLS model with fragment counts, which was the most appropriate model for prediction, was used to optimize the compositions. Subsequently, the authors identified the important substructures of monomers using least absolute shrinkage and selection operator (LASSO). The important substructures were used to select seed structures of monomers for structure generation. Another PLS model with distributed representation, called mol2vec, was constructed, because the ordinary fragment counts are unavailable for extrapolation. The PLS model estimated the polymer target property for screening novel structures. The major novelties of this study are to identify important substructures to the polymer target property and to apply mol2vec to design of network polymers. Eventually, we found a novel desired polymer through the composition optimization and demonstrated that virtual screening of monomers with distributed representation worked.