Multi-task deep learning for concurrent prediction of protein structural properties
Protein structural properties are diverse and have the characteristics of spatial hierarchy, such as secondary structures, solvent accessibility and backbone angles. Protein tertiary structures are formed in close association with these features. Separate prediction of these structural properties has been improved with the increasing number of samples of protein structures and with advances in machine learning techniques, but concurrent prediction of these tightly related structural features is more useful to understand the overall protein structure and functions. We introduce a multi-task deep learning method for concurrent prediction of protein secondary structures, solvent accessibility and backbone angles ({phi}, {psi}). The new method has main two deep network modules: the first one is designed as a DenseNet architecture a using bidirectional simplified GRU (GRU2) network, and the second module is designed as an updated Google Inception network. The new method is named CRRNN2. CRRNN2 is trained on 14,100 protein sequences and its prediction performance is evaluated by testing on public benchmark datasets: CB513, CASP10, CASP11, CASP12 and TS1199. Compared with state-of-the-art methods, CRRNN2 achieves similar, or better performance on the prediction of 3- and 8-state secondary structures, solvent accessibility and backbone angles ({phi}, {psi}). Online CRRN-N2 applications, datasets and standalone software are available at http://qianglab.scst.suda.edu.cn/crrnn2/.