Automating Water Orientation in Neutron Crystallography using Convolutional Neural Networks

Sammanfattning: One important aspect of determining the functionality of proteins is to find the direction of the dipole moment of surrounding water molecules. This is equivalent to finding the position of the hydrogen in water, however with the current atomistic model building tools this is a tedious and time-consuming task. The current main method for finding and building atomic resolution structures of biological macromolecules is known as crystallography. Using Neutron Crystallography an attempt at automating the task of determining the position of hydrogen and oxygen in water is made using 3D Convolutional Neural Networks. Seven networks are trained and evaluated, taking inspiration from current state of the art conventions and architectures (DenseNet and GoogLeNet). Using a volume as input with the target voxel in the center, each voxel of each water molecule is classified by the networks. A method to find the direction of the dipole moment with the network results is also formulated. The networks all perform similarly, despite varying depth and complexity. The Accuracy when finding hydrogen is around 60 percent and around 75 percent when finding oxygen, with the best networks having an AUC score of 0.631 and 0.851 respectively. A possibility to improve these results might lie in increasing the size of the volume input to the networks, since the surrounding dynamics of the crystal also affects the positioning of the hydrogen.