A key step in study of biosynthesis of membrane proteins is to look for the code that could be used to explain and predict which proteins would eventually be inserted in the membrane and which proteins would be secreted into the ER lumen when they cross the translocon channel. The aim of this work is to present an improvement of a previous method based on a local search approach. The proposed method relies on new in-depth biological observations to design a new search space for the local search algorithm. Experiments conducted on a dedicated dataset show that our new approach leads to improved outcomes in terms of prediction rates.