Feature Selection and Polydispersity Characterization for QSPR Modelling: Predicting a Tensile Property

CRAVERO, FIORELLA; SCHUSTIK, SANTIAGO; MARTÍNEZ, MARÍA JIMENA; BARRANCO, CARLOS D.; DIAZ, MÓNICA FÁTIMA; PONZONI, IGNACIO

Toledo

Conferencia; PACBB 2018: 12th International Conference on Practical Applications of Computational Biology & Bioinformatics; 2018

QSPR (Quantitative Structure-Property Relationship) models proposedin Polymer Informatics typically use reduced computational representations ofpolymers for avoiding the complex issues related with the polydispersion of theseindustrial materials. In this work, the aim is to assess the effect of this oversimplificationin the modelling decisions and to analyze strategies for addressing alternativecharacterizations of the materials that capture, at least partially, thepolydispersion phenomenon. In particular, a cheminformatic study for estimatinga tensile property of polymers is presented here. Four different computational representationsare analyzed in combination with several machine learning approachesfor selecting the most relevant molecular descriptors associated with thetarget property and for learning the corresponding QSPR models. The obtainedresults give insight aboutthe limitations of using oversimplified representations ofpolymers and contributewith alternative strategies for achieving more realisticmodels.