Protein secondary struture prediction using local adaptive techniques in training neural networks

Loading...
Thumbnail Image
Date
2006-06
Authors
Joseph, Annie Anak
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
One of the most significant problems in computer molecular biology today is how to predict a protein's three-dimensional structure from its one-dimensional amino acid sequence or generally called the protein folding problem, which leads to difficulties in determining the corresponding protein functions. Thus, this study involves protein secondary structure prediction using a neural network in order to solve the protein folding problem. The neural network used for protein secondary structure prediction is the Multilayer Perceptron (MLP) of the feed-forward variety. The network used consists of an input layer, a single hidden layer and an output layer. 120 proteins were taken from the Protein Data Bank (PDB) as training sets while 60 proteins were chosen randomly from the PDB as testing sets. Multiple sequence alignment (MSA) was used to obtain protein similar sequence while Position Specific Scoring matrix (PSSM) was used for network input. The hidden layer consists of 10 units while the output layer consists of 3 units. The result for secondary structure prediction are (I, 0, 0) for the Alpha helix, (0, 1 ,0) for the Beta Sheet and (0,0, 1) for the Coil. The training process of the neural network involves local adaptive techniques and dynamic adaptive techniques. Local adaptive techniques used in this study comprise of the Learning Rate by Sign Changes, Delta Bar Delta Rule, SuperSAB, Quickprop and RPROP while the dynamic adaptive techniques comprise of the Dynamic Momentum Factor (DMF), Dynamic Learning Rate 1 (DLRl) and Dynamic Learning Rate 2 (DLR2). From the simulation, the performance for Learning Rate by Sign Changes is Q3 = 86.67%, Delta Bar Delta Rule is Q3 = 85%, SuperSAB is Q3 = 88.33%, Quickprop is Q3 = 95% and Rprop is Q3 = 98.33% while for DMF, DLRl and DLR2 the performances are Q3 = 91.66%, Q3 = 93.33% and Q3 = 83.33% respectively. From the simulation, Rprop and Quickprop are superior compared to all other algorithms with respect to convergence time. However, the best result was obtained using the Rprop algorithm.
Description
Keywords
Struture prediction using , Training neural networks
Citation