Weights are adjusted so that the error between the observed output from each unit and the desired output specified by the target matrix is minimized. Each hidden unit transforms the signals received from the input layer by using a transfer function logsig to produce an output signal that is between and close to either 0 or 1, simulating the firing of a neuron. At each training cycle, the training sequences are presented to the network through the sliding window defined above, one residue at a time. The pattern recognition network uses the default Scaled Conjugate Gradient algorithm for training, but other algorithms are available (see the Deep Learning Toolbox documentation for a list of available functions). Then for each position in the window, we create an array of size 20, and we set the jth element to 1 if the residue in the given position has a numeric representation equal to j. In the following code, we first determine for each protein sequence all the possible subsequences corresponding to a sliding window of size W by creating a Hankel matrix, where the ith column represents the subsequence starting at the ith position in the original sequence. Thus, the input layer consists of R = 17x20 input units, i.e. In each group of 20 inputs, the element corresponding to the amino acid type in the given position is set to 1, while all other inputs are set to 0. Each window position is encoded using a binary array of size 20, having one element for each amino acid type.

We choose a window of size 17 based on the statistical correlation found between the secondary structure of a given residue position and the eight residues on either side of the prediction point. The input layer encodes a sliding window in each input amino acid sequence, and a prediction is made on the structural state of the central residue in the window.

For the current problem we define a neural network with one input layer, one hidden layer and one output layer.

Download secondary structure of protein