The alignments were also used to predict 2-state (exposed/buried) relative solvent accessible residues. Prediction accuracy was based on DSSP solvent accessibility definitions [52]. Relative accessibility was calculated by dividing the DSSP accessibility by the accessibility for a Gly-X-Gly tripeptide given by the method of Rose & Dworkin [53].
Three categories of relative accessibility were chosen for prediction: 25%, 5% and 0% accessible. Neural networks were trained with profiles from the HMMER2 profiles and the PSSM matrices from the PSIBLAST reports. The HMMER and PSIBLAST trained networks were also combined to give an average relative accessibility prediction.