Neural Networks for 3D Molecular Modeling

The image below illustrates why the features of neural network frameworks like Keras are so powerful for modeling based on 3D fields.

The most common use of neural network is for image analysis. In this context an image consists of a 2D grid of pixels, that has multiple data planes. These are often red, green, blue, and alpha transparency layers. The Keras toolset has a large library of capabilities to manipulate these layers together, or separately for recognizing objects in images, computing similarity and a host of other applications.

However the toolkits do not really depend on the origin of the data in the planes; they can perform the same convolutions, filtering, up-sampling, down-sampling and many other operations on whatever numbers are present.

The result is that if one is using the features for CoMFA or other 3D molecular analytics, there is no practical limit to the nature of data you can associate with the field. The original CoMFA used van der Waals and electrostatics fields. However the Keras framework allows one to add many other types - without increasing the complexity of the data structure. So one can easily add polarizability metrics, and electron densities specific to molecular orbitals to study if those features add predictive value to the model.

This is far more flexible that the original implementation as one can investigate the interactions among the fields.

The framework of python classes I'm developing to support this should be available on github in the coming months. It leverages the ubiquitous RDkit as well as tensorflow, numpy and other standard libraries.