This project built novel computational models using Spatial Transformer Convolutional Neural Networks (STNs) to better predict protein-ligand binding by recognizing perturbed three dimensional features. Models were constructed in both Caffe and Pytorch with the same architecture using the PDBbind dataset. The results show that the model in Caffe is able to predict the rigid body transformations for both translation and rotation of the ligand while the Pytorch model can learn translation but not rotation.

Authors:

David Ban (Presenter)
North Allegheny Senior High School

David Koes, University of Pittsburgh