Deep Residual Neural Networks for Protein Structure Prediction

Prediction of protein three-dimensional structure is a difficult computational biology problem with extensive applications to many areas, including pharmaceutical design and disease modeling. In the proposed work, a deep learning model is introduced to predict dihedral angles of the backbone (? and ?) and the distance matrices between the residues and physicochemical properties directly using sequence-derived features, physicochemical properties. The models are trained on the CASP7 subset of the ProteinNet dataset with taskspecific loss functions that are directly designed to regress angles and classify distances. The obtained system has an average root-mean-square deviation (RMSD) of 2.1 Å and coefficient of determination (R²) values of 0.39 for ? and 0.69 for ?, which indicates the strong predictions of structures. It is important to note that this design is lightweight and can be scaled and therefore implemented on the standard consumer-level computing hardware.