While many different aspects of retrieval algorithms (e.g., BLAST) have been studied in depth, the method for determining the retrieval threshold has not enjoyed the same attention. Furthermore, with genetic databases growing rapidly, the challenges of multiple testing are escalating. In order to improve search sensitivity, we propose the use of the false discovery rate (FDR) as the method to control the number of irrelevant ("false positive") sequences. In this paper, we introduce BLAST_(FDR), an extended version of BLAST that uses a FDR method for the threshold criterion. We evaluated five different multiple testing methods on a large training database and chose the best performing one, Benjamini-Hochberg, as the default for BLAST_(FDR). BLAST_(FDR) achieves 14.1% better retrieval performance than BLAST on a large (5,161 queries) test database and 26.8% better retrieval score for queries belonging to small superfamilies. Furthermore, BLAST_(FDR) retrieved only 0.27 irrelevant sequences per query compared to 7.44 for BLAST.
展开▼
