Protein secondary structure prediction using different encoding schemes and neural network architectures

Protein secondary structure prediction is very important for drug design, protein engineering and immunological studies. This research uses fully connected multilayer perceptron (MLP) neural network with one, two and three hidden layers to predict protein secondary structure. Orthogonal matrix, BLOSUM62 matrix and hydrophobicity matrix are used for input profiles. To increase the input information for neural networks, the combined matrix from BLOSLTM62 and orthogonal matrix and the combined matrix from BLOSUM62 and hydrophobicity matrix are also experimented. Binary classifiers indicate test accuracy of one hidden layer is better than that of two and three hidden layers. This may indicate that increasing complexity of architecture may not help neural network to recognize structural pattern of protein sequence more accurately. The results also show that the combined input profile of BLOSUM62 matrix and orthogonal matrix is the best one among five encoding schemes. While accuracy of the tertiary classifier reaches 63.20%, binary classifier for H/similar toH is 78.70%, which is comparable to other researchers' results.