Applications of deep neural networks to protein structure prediction
Protein secondary structure, backbone torsion angle and other secondary structure features can provide useful information for protein 3D structure prediction and protein functions. Deep learning offers a new opportunity to significantly improve prediction accuracy. In this dissertation, several new deep neural network architectures are proposed for protein secondary structure prediction: deep inception-inside-inception (Deep3I) networks and deep neighbor residual (DeepNRN) networks for secondary structure prediction; deep residual inception networks (DeepRIN) for backbone torsion angle prediction; deep dense inception networks (DeepDIN) for beta turn prediction; deep inception capsule networks (DeepICN) for gamma turn prediction. Every tool was then implemented as a standalone tool integrated into MUFold package and freely available to research community. A webserver called MUFold-SS-Angle is also developed for protein property prediction. The input feature to those deep neural networks is a carefully designed feature matrix corresponding to the primary amino acid sequence of a protein, which consists of a rich set of information derived from individual amino acid, as well as the context of the protein sequence. Specifically, the feature matrix is a composition of physio-chemical properties of amino acids, PSI-BLAST profile, HHBlits profile and/or predicted shape string. The deep architecture enables effective processing of local and global interactions between amino acids in making accurate prediction. In extensive experiments on multiple datasets, the proposed deep neural architectures outperformed the best existing methods and other deep neural networks significantly: The proposed DeepNRN achieved highest Q8 75.33, 72.9, 70.8 on CASP 10, 11, 12 higher than previous state-of-the-art DeepCNF-SS with 71.8, 72.3, and 69.76. The proposed MUFold-SS (Deep3I) achieved highest Q8 76.47, 74.51, 72.1 on CASP 10, 11, 12. Compared to the recently released state-of-the-art tool, SPIDER3, DeepRIN reduced the Psi angle prediction error by more than 5 degrees and the Phi angle prediction error by more than 2 degrees on average. DeepDIN outperformed significantly BetaTPred3 in both two-class and nine-class beta turn prediction on benchmark BT426 and BT6376. DeepICN is the first application of using capsule network to biological sequence analysis and outperformed all previous gamma-turn predictors on benchmark GT320.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 License