Quantitative assessment of protein function prediction programs

Fast prediction of protein function is essential for high-throughput sequencing analysis. Bioinformatic resources provide cheaper and faster techniques for function prediction and have helped to accelerate the process of protein sequence characterization. In this study, we assessed protein function prediction programs that accept amino acid sequences as input. We analyzed the classification, equality, and similarity between programs, and, additionally, compared program performance. The following programs were selected for our assessment: Blast2GO, InterProScan, PANTHER, Pfam, and ScanProsite. This selection was based on the high number of citations (over 500), fully automatic analysis, and the possibility of returning a single best classification per sequence. We tested these programs using 12 gold standard datasets from four different sources. The gold standard classification of the databases was based on expert analysis, the Protein Data Bank, or the Structure-Function Linkage Database. We found that the miss rate among the programs is globally over 50%. Furthermore, we observed little overlap in the correct predictions from each program. Therefore, a combination of multiple types of sources and methods, including experimental data, protein-protein interaction, and data mining, may be the best way to generate more reliable predictions and decrease the miss rate.