The underlying assumption of many automated sequence annotation methods is that similar sequences imply similar biological functions. The present paper re-examines this assumption. A novel measure based on a set of local BLAST alignments is introduced to define the overall similarity between two protein sequences. The relationships between yeast protein sequences and their biological functions in the context of gene ontology categories are presented, and the effects of the level of gene ontology terms and the size of gene ontology groups on the degree of similarity are studied. The similarity distributions at different levels of gene ontology trees are considered. To evaluate the theoretical prediction power of similar sequences, we compute the posterior probability of correct predictions. The results indicate that the posterior probability can serve as an important measure for automated protein function prediction.