A Study Of Protein Secondary Structure Prediction Methods Based On Neural Networks

Bioinformatics includes acquiring, processing, saving, distributing, analyzing and translating of biological information. It is a growing area of science that uses computational approaches to answer biological questions. By using mathematical, computer science and biological tools together, it clarifies and helps people in understanding the biological meaning of great volume of data. At present, its main research object is big biological molecule, and its main investigation tool is computer. With the development of Internet technology, biologists will have more opportunities to communicate with each other. Some departments provide services such as data sharing, inquiring and analyzing.At present, protein 3D structure prediction methods can be classified into two main kinds: molecular dynamics methods and knowledge based prediction methods. Molecular dynamics methods start from some basic principles and hypothesis. Knowledge based prediction methods start from observing and summarizing the rules of known protein structure.Proteins are polypeptide chains of molecular level, carrying out the most basic life functions. The physical and chemical properties of each of the 20 amino acids are fairly well understood. Polypeptide chains can fold into complicated 3D structure in correspondence to its function. The key step in predicting the folding of a protein is to predict its secondary structure. Secondary structures are made up of local folding structures, and they are often held by hydrogen bond.Protein secondary structure prediction is one of the important tasks of bioinformatics. As we have already seen with nucleotide sequences, all protein sequences, whether determined directly or through the translation of an open reading frame in a nucleotide sequence, contain intrinsic information of value in determining their structure or function. Protein secondary structure prediction forecasts the corresponding secondary structure of every amino acid in a sequence. It distinguishes helix, strand and non-routine structure. H, E and C represent helix, strand and non-routine structure respectively. The gaining upon ability, sorting ability and learning velocity of radial basis function neural networks are all better than BP networks. Aiming at solving this kind of complicated non-linear sorting problem, this paper analyzed the protein secondary structure prediction methods that are based on neural networks. Further, the paper discussed how to use radial basis function neural networks to predict protein secondary structure. The selection of data, the confirmation of network parameters and the influence of the parameters to network performance are also analyzed in this paper. The results of the model averaged about 69%, showing the feasibility and validity of the algorithm.Choosing effective training set is the basis of this research. Some sequences mainly consist of one type of the structures, so we must be careful when choosing the training set. To achieve higher prediction accuracy, the training set must be big enough and include all kinds of structures in proportion.Using the protein structure data provided by PDB, the database used by this paper is constructed. Constructing a database for this study is very important because of the following two points. First, the records in PDB are in detail, so the corresponding part must be extracted. Second, since the training and testing sets must be big enough, a number of proteins must be chosen from PDB.The correlation between neighboring residues provides important information in secondary structure prediction. Since it lacks the understanding of the correlation, the prediction accuracy of simple RBF neural networks is limited. In order to take more consideration of this kind of correlation, second level RBF neural networks is introduced.Using evolution information can improve the prediction accuracy of neural networks. Multiple alignment can produce position specific profiles. They can describe which residues can be exchanged against which others at which position,...