The process of assessing the prediction methods resulted in different neural networks that were trained with different alignment data. Each of the networks were combined and the average taken for each predicted state, be it helix strand or coil.
The outputs from each of the networks trained previously, were also taken and positions examined where the predicted state was identical in all methods. For positions where there was a `jury agreement' (identical predictions by all methods), the average Q3 accuracy was 82%. Residues where the predictions did not all agree were classified as a `no jury' positions (see Figure 3). Positions where there was a full agreement in the predicted state between the different neural network methods were taken as the final prediction. Positions where there was `no jury' were used to train a separate neural network. The final prediction was obtained by replacing the original `no jury' positions with the predictions from this network.