Can we use the Gene Ontology database as a gold standard to test the performance of PSI-BLAST?

Question

[radinms]

I have an assignment question about this. We created a ROC plot to compare PSI-BLAST results agains GO database and looked for how many true and false positives hits were present in our PSI-BLAST results. I think the answer is no. We can't use the GO database as a 'gold standard' since the essence of what PSI-BLAST does is so different from GO, but we could use it to double check and point out certain false positive in PSI-BLAST results? I found a few papers about increasing PSI-BLAST accuracy and none mention GO, however they do mention using literature and even google to increase accuracy so we should be able use GO database for this purpose as well? Could someone elaborate with a comprehensive explanation? Thanks in advance!

[Jean-Karim Heriche]

It depends a bit on what you want to predict with PSI-BLAST because some subsets of the ontology may be more easily inferred from sequence than others. In any case if you want to use GO annotations, then you should only use those that have been manually curated and assigned an experimental evidence code because most of the annotations are transferred based on sequence similarity and alignments.

Answer 1

[Mensur Dlakic]

Not all levels of GO are based on sequence relationships (homology) between proteins. Some GO levels are based on functional relationships that can exist in the absence of homology, and (PSI-)BLAST has nothing to detect in such instances. Generally speaking, benchmarking PSI-BLAST based on GO will work only for GO classification levels that are homology-based.

[radinms]

Yes, that's what I thought as well. Thank you for the clarification.