(346r) Discovery of Chemical Descriptors/Features in Predicting Protein-Ligand Binding Properties Using 3D Convolutional Neural Network

Authors: 
Ashraf, C., University of Washington, Seattle
Beck, D., University of Washington
Pfaendtner, J., University of Washington
Structure-based virtual screening techniques are some of the most successful methods for enhancing various processes of interest such as artificial dialysis, drug delivery, purification of specialty chemicals, sensing and detection, and waste-water treatment. With the growing needs for advanced materials that facilitate low energy separation, capture or delivery of molecules with high specificity, there has been a significant effort by the researchers to apply advanced Machine Learning (ML) models to estimate material properties and design novel materials. One such application is to predict protein-ligand binding pockets and/or affinities using 3D Convolutional Neural Networks (CNN). Here, the 3D protein/ligand structure is treated as a 3D image and pharmacophore properties are used as the features/channels for the CNN. This approach suffers from two basic limitations: the features generated should be rotationally and translationally invariant and only the relevant features should be used to train the network. To eliminate these limitations, in this study, we use an Euclidean Neural Network,1 which is a CNN with specially designed convolution kernel that transforms the input molecular features such that they are rotationally and translationally invariant and therefore leading to more robust learning of the network. At the same time, we aim to incorporate feature extraction directly into the ML model. In such an approach, the CNN described above will become the first component of the model, which is then trained together with a predictive component to extract features that are useful in solving a specific task such as predicting protein-ligand binding properties. With this learn-test-design approach, it is possible to discover specific features/descriptors and identify relationships relevant to the specific objective, which will lead to the design of novel materials for particular applications.

(1) Thomas, N.; Smidt, T.; Kearnes, S.; Yang, L.; Li, L.; Kohlhoff, K.; Riley, P. Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for 3D Point Clouds. 2018.