Protein Residue Contact Prediction And Its Application In Protein Structure Prediction

In protein native structures,amino acid residue contacts encompass abundant protein structure information.With the accumulation of protein sequences,the advancement of protein co-evolution algorithms,and the development of machine learning,protein residue contact prediction has gradually become one of the most important methods to predict the3D structure of proteins.With the use of residue contact prediction,the complexity of protein structures prediction and the structure search space could be greatly reduced.The efficiency and accuracy of protein structure prediction could also be improved significantly.Nowadays,most residue contact algorithms treat the co-evolution score matrix as images.Then the mature image processing algorithms will be applied to predict the protein residue contact.These algorithms have exhibited satisfactory performance.How-ever,most algorithms only utilize limited biology knowledge.How to integrate biology knowledge into contact prediction has become an important direction in this field.In this study,we proposed two algorithms,RDb₂C and Amoeba Contact.These algorithms were designed to predict the?residues contacts and the general residue contacts,respectively.RDb₂C used the ridge detection in image processing to extract the band-like signals in contact maps.Then,we constructed the random forest model to predict the?residues contacts.RDb₂C outperformed bbcontacts,the best algorithm in this field,in 2 well-accepted dataset?Beta Sheet916 and Beta Sheet1452?and could improve the?protein structure prediction.The Amoeba Contact algorithm were designed to predict general residue contacts.We introduced two new normalization operations in Amoeba Contact.These new operations could integrate the sparseness constrains of residue contact into deep learning frame-works,hence lead to performance improvement.We also utilized the automated neural architecture search technology to find an optimal network architecture for protein residue contact prediction.By generalizing the Amoeba Contact to different contact cutoffs,more delicate contact constraints could be obtained.Based on these constraints,we designed a gradient descent-based algorithm GDFold to predict the structure of proteins.GDFold has similar performance to the current advanced algorithms like Raptor X-Contact.