Research On Protein Contact Map Prediction Based On Deep Convolutional Neural Network

Protein structure information plays an enormous role in many fields including biology,medicine,pharmacology.Nevertheless,in quantitative terms,there is a big gap between structures obtained by wet-lab experiments and exist in nature.Thus de novo protein structure prediction has significant theoretical and practical value.Protein contact-map provides topology constraint and thus can be an effective help for the reconstruction of protein structure.There are differences between contact map prediction and classic classification problems,i.e.,the human-defined threshold value of being in contact,the inaccurate coordinates of atoms in a solved structure and the imbalanced training data.So certain adjustment should be made if we want to apply traditional machine learning algorithms to the problem.Furthermore,we focus on the weakness of current methods.Our experiment results show that state-of-the-art methods still show deficiencies in the contact prediction of proteins with low-homology information.(1)In this thesis,we first propose a differentiated contact map prediction method.We build a deep convolutional neural network and take predicted three-state secondary structure,predicted solvent accessibility,direct co-evolution information,mutual information,residual pair potential energy as input features.Furthermore,we choose the focal loss,which is more sensitive to wrong predictions,over cross-entropy loss.Sufficient comparative experiments show that our method's advantages.(2)We propose a brand-new pipeline which doesn't rely on homology information.It makes prediction only using the information coming from the target itself.According to our experiments,the proposed method is superior to other methods in the prediction of lowhomology targets,providing a new way for contact-map prediction.