A nuclear export signal (NES) is a protein localization signal, which is involved in binding of cargo proteins to nuclear export receptor, thus contributes to regulate localization of cellular proteins. Consensus sequences of NES have been used to detect NES from protein sequences, but su®er from poor predictive power. Some recent peering works were proposed to use biochemical properties of experimental veri¯ed NES to re¯ne NES candidates. Those methods can achieve high prediction rates, but their execution time will become unacceptable for large-scale NES searching if too much properties are involved. In this work, we developed a novel computational approach, named NES-REBS, to search NES from protein sequences, where biochemical properties of experimental veri¯ed NES, including secondary structure and surface accessibility, are utilized to re¯ne NES candidates obtained by matching popular consensus sequences. We test our method by searching 262 experimental veri¯ed NES from 221 NEScontaining protein sequences. It is obtained that NES-REBS runs in 2-3 mins and performs well by achieving precision rate 47.2% and sensitivity 54.6%.