Multi-instance learning for structure-activity modeling for molecular properties

Abstract

In this paper, the approach of multi-instance learning is used for modeling the biological properties of molecules. We have proposed two approaches for the implementation of multi-instance learning. Both approaches are based on the idea of representing the features describing the molecule as a one vector, which is produced from different representations (instances) of the molecule. Models based on the approach of multi-instance learning were compared with classical modeling methods. Also, it is shown that in some cases, the approach of multi-instance learning allows to achieve greater accuracy in predicting the properties of molecules.