Strategy of Structure Generation within Applicability Domains with One-Class Support Vector Machine
Hiromasa Kaneko, Kimito Funatsu · Bulletin of the Chemical Society of Japan · 2015
Abstract Generation of virtual chemical structures is applied in material, product, and drug designs to obtain structures having desired activity or properties. Quantitative structure–activity relationship (QSAR) models and quantitative structure–property relationship (QSPR) models are used to estimate values of activity and those of properties of structures, respectively. However, estimated values are unreliable when new structures are out of an applicability domain (AD) which is defined using a training data set. Data density around a new structure, which can handle nonlinearities between descriptors and multimodal data distributions, can be an index of ADs. We focus on one-class support vector machine (OCSVM) as a data density estimation method and propose a strategy of structure generation within ADs. The partial derivative of an OCSVM model with respect to each descriptor is used as a guideline to change each descriptor to get structures within ADs. It was confirmed that structures could be changed to increase data density around the structures, and could be generated within ADs through QSAR and QSPR data analyses.