Fast, conformation-independent charges with graph convolutional networks.

Abstract

Two key disadvantages make charge models based on quantum chemistry (QC) calculations difficult to apply consistently across biomolecular simulations. Firstly, QC calculations become intractably slow for large molecules such as proteins. Secondly, QC properties are typically dependent on the coordinates supplied and the charges generated therefore vary between different conformations. Neural network models such as graph convolutional networks offer solutions that do not require QC calculations, remaining conformer-independent and scaling efficiently.