Protein structure featurization via standard image classification neural networks

Many applications in the biomedical domain involve the detailed molecular and functional characterization of macro-molecules such as proteins. Where possible, this involves the knowledge of detailed 3D coordinates of every atom within a protein. At the same time, machine learning has become the basis of much innovation within this domain in recent years. There are, however, a few challenges in applying machine learning to 3D protein structures, such as variability in size and high dimensionality of the data. It would therefore be beneficial to be able to map every protein structure to a smaller fixed-dimensional representation that is directly learned from the structure without manual curation. In addition, it would be valuable for biomedical researchers if such approaches would require little method development and instead draw from cutting-edge research such as image classification via deep neural networks. Here, such an approach is outlined that first re-formats protein structures as 2D color images and then applies off-the-shelf neural networks for image classification. It is shown that such neural networks can be trained to effectively encode the CATH protein classification database and that feature vectors extracted from such networks, once trained, can be transferred to a completely new task that is likely to benefit from molecular protein information, namely that of small molecule binding.