[dekhn]

Yep! In my work we had a golden labelled set of known-homologous proteins, as well as sets of unknown-relationship and known-not-to-be-homologous (based on superfamily). We blasted all-vs-all and after dealing with a number of BLAST issues finally demonstrated that the searches produce false positives at the expected rates. I gained a lot of respect for the folks who implemented the statistical framework behind blast E-values.

I had another advisor who thought they were good at homology, so they'd take protein matches with an E-value of 10 (almost certainly random chance) and munge the data to make up a story. Most of my work was about using very stringent blast searches to make profiles, but all that work is superseded by HMMER now.