Abstract:
One key method of automatic functional annotation of a prokaryote gene is
finding BLAST hits to the gene in question that have functional annotations, choosing the
best single hit, and copying the annotation from that hit if it is of sufficient quality as
measured by a p-value or other criterion. In the JCVI prokaryote automatic functional
annotation system, the best hit is chosen by looking up categories in a manually constructed table stating how reliable the annotation is depending on who made the
annotation, what percent identity the BLAST hit had, and what percentage of the query
gene and the hit gene were involved in the match.
Constructing this table is labor-intensive; and humans are incapable of processing
enough data to construct it correctly. I therefore reduced the data requirements by
breaking the table into orthogonal components; and I developed an iterative method to minimize the least-squares error of the table on a training set. I also constructed a
validation set of 50,000 manually-annotated proteins from JCVI data, and developed a
protein name thesaurus and ontology to make it possible to tell when two names meant
the same thing, or when one name was a more-specific refinement of another name.
Training on 9/10 of the validation set, and testing on the held-out 1/10, showed an
improvement in accuracy from 61.5% to 72.4%.
