3D Deformable Protein Shapes Classification based on Triangles-Stars and Composite Deep Neural Networks

Abstract

In this paper, the problem of 3D protein deformable shape classification is addressed. Proteins are macromolecules with deformable shapes, their classification based only on their molecular surfaces is challenging problem. In addition, the protein shapes are related to their functions which makes their classification an important task. Triangular meshes (graphs) my be considered to represent the protein molecular surfaces. In this paper, we propose a new deep learning-based approach for 3D protein deformable shape classification. We propose a 3D deformable shape descriptor and a composite deep neural network. The shape descriptor is based on a set of global and local features resulting from the decomposition of protein 3D shapes into triangles-stars, following the different neighborhood order considered. The classification is performed by a composite deep neural network, each branch corresponding to a neighborhood order. The proposed approach is evaluated against 3D protein benchmark repositories and state-of-the-art methods. Our experimental results show and demonstrate the high performances and the effectiveness of our approach.