Research On Transmembrane Helices Prediction For Membrane Protein

Recent studies have shown that the structural or functional changes of some membrane proteins have impact to the human diseases. In addition, the receptor membrane proteins have also become important targets for drug design. Thus, revealing the structure of membrane proteins has great significance. Experimentally determining the structure of membrane proteins is a complex and time-consuming task, and predicting the structure of membrane protein by using pattern recognition technology has become a promising route.This paper focuses on the transmembrane helices prediction. Self-organizing map (SOM) is utilized to learn sample distribution knowledge. The learned knowledge is encoded in the codebook vectors of the SOM. These codebook vectors are used as the optimized training samples. In the prediction stage, K-NN and support vector machine (SVM) are taken as the classifiers:the probability of being a transmembrane helix (TMH) for each amino acid is firstly computed; then the dynamic segmentation method is applied to obtain the final TMHs.Experimental results on two benchmark datasets with cross-validation shown that the proposed method significantly improves the prediction speed and the also helps to improving the prediction accuracy. We also compared the proposed method with 8 popular existing TMH predictors, such as THUMBUP, HMMTOP, SOSUI, DAS-Tmfilter, TOP-PRED, TMHMM, PHOBIUS, and MemBrain. It is found that the proposed method is only slightly inferior to MemBrain and acts as the second-best performer.