Fibronectin type III domains in yeast detected by a hidden Markov model

relies on all proteins from an organism being in sequence databases. In this manner, if only one protein within a given pI and mass range is found with a certain aminoor carboxy-terminal sequence tag, one can be confident that there is no other, as yet undescribed, protein that could otherwise match the tag. In fully sequenced organisms, the procedure is thus self-checking. The specificity of sequence tags may be an issue in larger organisms: whereas there are (for example) 3 200 000 combinations of five amino-acid tags, protein amino termini have biased sequences and many amino termini are shared. However, protein carboxyl termini have almost random sequences (data not shown) so their sequence tags should be more specific. Other factors to consider will be the accuracy of sequence data that can be obtained from proteins purified from twodimensional gels, and the accuracy of prediction of protein open reading frames in genome/proteome databases. Large-scale protein characterization projects will define the effect of these factors and thus the utility of sequence tags for protein identification.