OXBench: a benchmark for evaluation of protein multiple sequence alignment accuracy.

The alignment of two or more protein sequences provides a powerful guide in the prediction of the protein structure and in identifying key functional residues, however, the utility of any prediction is completely dependent on the accuracy of the alignment. In this paper we describe a suite of reference alignments derived from the comparison of protein three-dimensional structures together with evaluation measures and software that allow automatically generated alignments to be benchmarked. We test the OXBench benchmark suite on alignments generated by the AMPS multiple alignment method, then apply the suite to compare eight different multiple alignment algorithms. The benchmark shows the current state-of-the art for alignment accuracy and provides a baseline against which new alignment algorithms may be judged. The simple hierarchical multiple alignment algorithm, AMPS, performed as well as or better than more modern methods such as CLUSTALW once the PAM250 pair-score matrix was replaced by a BLOSUM series matrix. AMPS gave an accuracy in Structurally Conserved Regions (SCRs) of 89.9% over a set of 672 alignments. The T-COFFEE method on a data set of families with <8 sequences gave 91.4% accuracy, significantly better than CLUSTALW (88.9%) and all other methods considered here. The complete suite is available from http://www.compbio.dundee.ac.uk. The OXBench suite of reference alignments, evaluation software and results database provide a convenient method to assess progress in sequence alignment techniques. Evaluation measures that were dependent on comparison to a reference alignment were found to give good discrimination between methods. The STAMP Sc Score which is independent of a reference alignment also gave good discrimination. Application of OXBench in this paper shows that with the exception of T-COFFEE, the majority of the improvement in alignment accuracy seen since 1985 stems from improved pair-score matrices rather than algorithmic refinements. The maximum theoretical alignment accuracy obtained by pooling results over all methods was 94.5% with 52.5% accuracy for alignments in the 0-10 percentage identity range. This suggests that further improvements in accuracy will be possible in the future.