We begin with two assumptions: 1) proteins with a similar function tend to be closely related in some feature space - e.g. structure, sequence, evolution, etc.; and 2) nodes in the Gene Ontology (GO) are a reasonable representation of protein function. We, therefore, have developed a method for using nearest neighbors in feature space to predict protein function in the form of a set of one or more GO nodes. The example we present uses the following strategy: 1) identify close neighbors of a target protein sequence using PSI-BLAST; 2) collect the GO nodes associated with these neighbors using a curated mapping from Swiss-Prot to GO, weighting each member of the collection relative to the PSI-BLAST evalue; 3) categorize the collection of GO nodes based on their distribution in the Gene Ontology structure, utilizing a technology called the Gene Ontology Categorizer (GOC). To establish appropriate parameter settings for GOC and evaluate our ability to predict protein function, we analyzed a test set of protein sequences with known GO annotation.