Search

Protein structure prediction is a main task in the field of bioinformatics, and the prediction of protein secondary structure is the key point of this task. Extracting representative features and efficient classification methods are the basis of the prediction process. In this paper, a prediction method based on BP neural network is proposed. We use valid protein features extracted by variable-sized sliding window and two different encoding modes (5-bit encoding and Profile encoding) as input data, to make predictions for the secondary structure of proteins. The prediction accuracies are calculated by Jackknife test on three commonly used low-similarity protein datasets: 25PDB, 1189 and 640, and this method achieves a high overall accuracy upon these three datasets.

In order to filter out the PPI (protein-protein interaction) data noise and improve the accuracy of protein complexes prediction, this paper proposed a method to integrate biological information and topology information as the weight of the PPI network. There are also some traditional methods to deal with the data noise by topology information of biological information, but fewer of them integrate those two aspect. In this paper, we introduced a method to extract the feature vectors for protein couples or the interactions instead of single proteins based on both topology and biology information, then the cosine theory is used to measure the weight of PPI network. At last, it used four classical predicting protein complexes algorithms to detect protein complexes from different weighted PPI network. The result shows that our method performs better both on filtering out data noise and predicting protein complexes.