ESG: Extended Similarity Group method for automated protein function prediction

We present here the Extended Similarity Group (ESG) method, which annotates query sequences with Gene Ontology (GO) terms by assigning probability to each annotation computed based on iterative PSI-BLAST searches. Conventionally sequence homology based function annotation methods, such as BLAST, retrieve function information from top hits with a significant score (E-values). In contrast, the PFP method, which we have presented previously, goes one step ahead in utilizing a PSI-BLAST result by considering very weak hits even an E-value of up to 100 and also by incorporating the functional association between GO terms (FAM matrix) computed using term co-occurrence frequencies in the UniProt database. PFP is very successful which is evidenced by the top rank in the function prediction category in CASP7 competition. Our new approach, ESG method, further improves the accuracy of PFP by essentially employing PFP in an iterative fashion. An advantage of ESG is that it is built in a rigorous statistical framework: Unlike PFP method that assigns a weighted score to each GO term, ESG assigns a probability based on weights computed using the E-value of each hit sequence on the path between the original query sequence and the current hit sequence