Motivation: The prediction reliability of seven multiple alignment servers currently available on the Internet (ClustalW, MAP, PIMA, Block Maker, MSA, MEME and Match-Box) has been evaluated in terms of power (sensitivity) and confidence (selectivity). Therefore, the alignments obtained have been respectively compared to refined structural alignments for 20 families of related proteins with low levels of identity.
Results: Results clearly show that any powerful method remains reliable when the rate of identity falls. For some methods, power and confidence decrease linearly with the rate of identity, while other methods emphasize reliability at the cost of a lower power. Increasing the number of related sequences included in the alignment may either improve or decrease the quality of the predictions substantially. For some methods, the gain in power or in confidence is quite systematic; for others, the effect of the addition of homologous sequences is highly unpredictable. Extracting the consensus between two different methods may increase the overall confidence of the predictions tremendously. Our conclusions induce users of sequence alignment methods on the Internet to select the most suitable technique according to their requirements in terms of selectivity and sensitivity.
Availability: The aligned sequences of the 20 alignments of structure can be obtained automatically by sending the message 'send: cabios_tests.txt' by e-mail to 'matchbox@biq.fundp.ac.be'.
Contact: eric.depiereux@fundp.ac.be