2017.12.28 Protein Contact Prediction and Folding by Deep Learning
题 目: Protein Contact Prediction and Folding by Deep Learning
报告人: Professor Jinbo Xu
Toyota Technological Institute
The University of Chicago
时 间: 2017年12月28日(周四)15:30
地 点: 北京大学化学与分子工程学院A204报告厅
Ab initio folding is one of the most challenging problems in Computational Biology. Recently contact-assisted folding has made some progress on this problem, but it requires accurate inter-residue contact prediction, which by existing methods can only be achieved on some proteins with a very large number of sequence homologs. To deal with proteins without many sequence homologs, we have developed a novel CASP-winning deep learning (DL) method for contact prediction that formulates it similarly as image semantic segmentation and then applies the concatenation of two deep residual neural networks (ResNet). The first ResNet conducts convolutional transformation of 1-dimensional protein features to capture sequential context of one residue and the second conducts convolutional transformation of 2-dimensional features to exploit higher-order residue correlation. Experimental results suggest that our DL method doubles the accuracy of pure co-evolutionary methods on proteins without many sequence homologs and can fold many more proteins than ever before. Our method is officially ranked No. 1 for contact prediction in the latest protein structure prediction competition (CASP12) and also works well on membrane proteins and inter-protein contact prediction even if trained by single-chain non-membrane proteins.
See http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005324 and http://www.cell.com/cell-systems/fulltext/S2405-4712(17)30389-7 for technical details.