The effect on accuracy of alternative 8 to 3 state reductions

DSSP[38] provides an 8 state assignment of secondary structure denoted by single letter codes: H ($\alpha$-helix), T ($\beta$-turn), S (bend), I ($\pi$ helix) , G ( ${\rm 3_{10}}$-helix), E ($\beta$-strand), B ($\beta$-bridge) and C (not HTSIGE or B). However, prediction methods are normally trained and assessed for only 3 states (H,C,E), so the 8 states must be reduced to 3. Here we consider the effect on accuracy of applying three different published 8 to 3 state reduction methods when testing.

• Method A: E and B to E, G and H to H. Rest to coil[27]

• Method B: E as E, H as H. Rest to coil including EE and HHHH[28]

• Method C: GGGHHHH redefined as HHHHHHH, then B, and GGG to coil, with H to H and E to E[30]

Method A treats both isolated $\beta$-bridges and residues that are part of a $\beta$-sheet as 'extended'. ${\rm 3_{10}}$-helix and $\alpha$-helix are treated as 'helix', and all other states treated as 'coil'. Method B translates more secondary structures into coil. This includes all ${\rm 3_{10}}$-helix, $\alpha$-helix that is a single turn, isolated $\beta$-bridges and $\beta$-strands that are only two residues long. The rationale for this reduction is that short secondary structures are normally of marginal stability and also variable within protein families. Table 10 shows reduction Methods B and C on average to contain 4% more coil, 3% less helix and 2% less strand than Method A.

Table 11 summarises the difference in Q₃ accuracy obtained by varying the reduction of 8 state DSSP definition to 3 states. The original value for the consensus method was 74.8% (Method A). This increases to 77.9% for Method B. Table 11 shows that reducing both single strand 'B' to coil and ${\rm 3_{10}}$-helix 'G' to coil, gives the greatest stepwise increase (2.7%).

Table 12 shows that all prediction methods appear to improve in accuracy with comparison to Method A, when one uses Method B, as the method for the 8-state reduction. The apparent improvement is between 2.2 and 4.9%. The PREDATOR[28] method improves by the largest extent (4.9%), as a reflection of PREDATOR being trained with Method B as the reduction scheme. STRIDE [40] values have also been included in Table 12 for comparison. Prior to this study, PHD[64] and DSC[29] used the same reduction method (Method A), but NNSSP[30] and PREDATOR[58,28] used methods B and C respectively. Unless the same reduction method is used, an objective comparison can not be made.