Assessment of secondary structure prediction accuracy

The accuracy of prediction was assessed by comparison to DSSP [52]. We applied the same 8 to 3 state reduction scheme throughout. E and B were taken as strand, H as helix and all other states were taken as coil. This removes the 3¹⁰-helix state from the definition, but given that the 3¹⁰-helix state represents a weak 1 kcal/mol hydrogen bond, this seems reasonable, as it does not normally represent core secondary structure. Single residue $\beta$-bridges (B) were included, as they do form part of core structures, however these structures are typically difficult to predict accurately. Removing single residue $\beta$-bridges improves the apparent average Q₃ prediction accuracy by at least 1% (data not shown).