Improving protein fold recognition by extracting fold-specific features from predicted residue–residue contacts

• DOI: 10.1093/bioinformatics/btx514
• Publication Date: 2017-08-09T19:13:47Z
• Authors: Jianwei Zhu; Haicang Zhang; Shuai Cheng Li; Chao Wang; Lupeng Kong; Shiwei Sun; Wei-Mou Zheng; Dongbo Bu
• Keywords: Template; Residue (chemistry)

Accurate recognition of protein fold types is a key step for template-based prediction structures. The existing approaches to mainly exploit the features derived from alignments query against templates. These have been shown be successful at family level, but usually failed superfamily/fold levels. To overcome this limitation, one points explore more structurally informative proteins. Although residue-residue contacts carry abundant structural information, how thoroughly these information still remains challenge.In study, we present an approach (called DeepFR) improve basic idea our extract fold-specific predicted proteins using deep convolutional neural network (DCNN) technique. Based on features, calculated similarity between and templates, then assigned with type most similar template. DCNN has showed excellent performance in image feature extraction recognition; rational underlying application that contact likelihood maps are essentially analogy images, as they both display compositional hierarchy. Experimental results LINDAHL dataset suggest even extracted alone, achieved success rate comparable state-of-the-art approaches. When further combining traditional alignment-related increased 92.3%, 82.5% 78.8% family, superfamily levels, respectively, which about 18% higher than 6% level 1% level. An independent assessment SCOP_TEST consistent improvement, indicating robustness approach. Furthermore, bi-clustering compatible hierarchy proteins, implying fold-specific. Together, orthogonal combination them could greatly facilitate levels structures.Source code DeepFR freely available through https://github.com/zhujianwei31415/deepfr, web server http://protein.ict.ac.cn/[e-mail] or [e-mail] data Bioinformatics online.