Algorithms Of Fragment Library Construction And Their Application In Protein Structure Prediction

As one of the most important and challenging problem,ab initio protein structure prediction is the process,which folds proteins from random conformations to their native structures.The research of ab initio protein structure prediction is essential to under-standing the protein folding mechanism and detecting the relationship between protein conformational changes and their functions.Effective conformational searching algo-rithms and accurate potentials are the main components of protein structure prediction and thus guarantee the high accuracy of predicted models.Exemplified by the great suc-cess of Rosetta,fragment assembly has become the most popular conformational search algorithm over the past two decades.Therefore,the quality of fragment library determines the efficiency of conformational search and the accuracy of protein structure prediction.Over the last decade,on one hand,great progress has been made on machine learn-ing field,especially on the deep learning techniques such as recurrent neural networks and convolutional neural networks.On the other hand,besides the secondary structure prediction,local structure predictions such as contact prediction and torsion angle predic-tion have been developed rapidly.The novel deep learning techniques and high-accuracy protein local structure predictions pave the way for fragment library construction.In this work,we first designed LRFragLib based on logistic regression models.By utilizing the features of primary sequences,physicochemical properties and secondary structures,LRFragLib searches 7-10 residue fragments for target proteins.Compared with state-of-the-art algorithms,LRFragLib improved the proportion of near-native frag-ments by a large margin and predicted more accurate protein models by combining to ab initio protein structure prediction software.Subsequently,we designed FragMove,an substitution pattern of dihedral angles based on the fragment libraries generated by LRFragLib and incorporated it into REMC.As a consequence,the accuracies of sec-ondary structures and tertiary structures as well as the folding efficiency improved.In the third step,we designed DeepFragLib for identifying 7-15 residue fragments by the cutting-edge deep learning techniques.DeepFragLib consists of the classification mod-ule with bidirectional LSTM layers after knowledge distillation,the regression module with ResNeXt architecture containing self-designed cyclically dilated convolution layers and the fragment selection module.Systematic analysis demonstrated that DeepFragLib outperformed LRFragLib and the other algorithms by a large margin.Moreover,the eval-uation with Rosetta suggested that DeepFragLib in deed contributed to ab initio protein structure prediction.