Prediction of Protein Secondary Structure Based on a Multi-modal Neural Network: with Modified Profiles of MSA and PSSM

概要

論文の詳細を見る
Prediction of protein secondary structure is considered as an important step towards elucidatingits three-dimensional structure, as well as its function. We have developed a multi-modal neuralnetwork for predicting protein secondary structure. The prediction is based on the frequency profile ofmultiple sequences alignment and the position specific scoring matrices (PSSM) generated byBLOCK. The multi-modal neural network is composed of two steps: The first step is to develop threeneural networks to predict the secondary structure states of proteins: α-helix, β-sheet andnon-regular structure respectively. The single-state prediction neural networks use a local inputwindow of consecutive amino acids to predict the secondary structure state of the amino acid locatedat the center of the input window; The second step is to develop a decision neural network to combineall of the single-state predictions to obtain an overall prediction on three states. This method gives anoverall accuracy of 67.8% when using seven-fold cross-validation on a database of 126non-homologous proteins. To improve the accuracy further, majority decision is introduced to eachnetwork for single-state prediction in the first step. By using majority decision, the overall accuracy isimproved to 70.2% with corresponding Matthews' correlation coefficients Cα =0.61, Cβ=0.48.
宮崎大学の論文