Conclusions

In this paper, the effect of training a two-level neural network algorithm for protein secondary structure prediction with the same sequences presented as different alignment profiles has been investigated. The general conclusions are:

1.

By appropriate selection of database searching method, alignment algorithm and scoring scheme, the prediction accuracy for the same sequences using the same basic algorithm is improved by 7% points from 69.5% to 76.4%.

2.

Solvent accessibility prediction accuracy has been improved by 1.2% to 76.2% for a two state model, and has been extended to include specific prediction of the 25, 5 and 0% relative accessibility states.

3.

Confidence in prediction has been improved. Residues predicted with a confidence of 5 and greater, will be on average 84% accurate and cover 68% of residues. The average prediction accuracy per protein is 76.4% with a standard deviation of 8.4%.

4.

Optimisation of learning parameters and training algorithm, may further extend the accuracy of the secondary structure prediction method.

5.

In the years from 1993 to 1999, prediction accuracy has improved from 70.6% [17] to over 76% (this work). Most of this improvement has come from more sophisticated use of sequence alignments rather than enhancements to the neural network algorithm.

6.

The most dramatic improvements in prediction accuracy have come from the use of PSIBLAST position specific scoring profiles in preference to profiles derived from global multiple alignment methods such as CLUSTALW and AMPS.

Given the expansion of structural genomics projects, which aim to solve protein structures much more rapidly, the exploitation of these data will only extend the ability to predict protein structure ever more accurately.