In this work, the SNNS neural network package from Stuttgart University was applied. SNNS allows for rapid prototyping of neural networks, while also allowing incorporation of the resulting networks into a ANSI C function for use in stand-alone code.
The network ensemble consisted of 2 artificial neural networks. The first was a network with a sliding window of 17 residues over each amino acid in the alignment, plus the addition of a conservation number as the input nodes. The network then comprised a further 9 hidden nodes and 3 output nodes. The output from this network was windowed into 19 residues, plus a conservation number , which formed the input to the second network. This second network also had 9 hidden nodes and 3 output nodes. 250 epochs of Scaled Conjugate Gradient (SGC) training  were applied, from an initial random weighting of node values of between 0.005 and -0.005. No optimisation was carried out for the number of training cycles.
For the PSIBLAST profiles and the HMMER profiles, only the windowed values for the profiles were applied, no conservation number was added.