Evaluating GO-based Semantic Similarity Measures

Motivation: While several efforts have been made in measuring GO-based protein semantic similarity, it is still unclear which are the best approaches to measure it and furthermore whether electronic annotations should be used. Results: We studied the behaviour of 8 distinct semantic similarity measures as function of sequence similarity with and without electronic annotations. We found that 5 of these measures shared a cumulative normal distribution pattern, which is likely inherent to the relation between functional and sequence similarity. We also present a novel graph-based measure for protein semantic similarity, which produced bet-ter results than the other measures studied