Predicting Protein Submitochondria Locations Based On Multi-Features Fusion

With the success of human genome project, abundance of unknown functional proteins appear in the database and the most important task in today is to analyzing these unknown functional proteins. Researchers begin to interest in subcellular organelle location for analyzing the functions of protein when the subcellular location as analyzing methods has achieved certain standard. Due to the experiment approach is time-consuming and expensive, thus the computational methods for predicting the protein subcellular organelle location has become the current research focuses.In this dissertation, we systematically studied the protein submitochondria location from the aspects of the dataset construction, the extraction and optimization of feature vectors, establishing the prediction algorithm and the generalization of algorithm. The main research findings are as follows:1. Because there are a few sequences in the current dataset of submitochondria location constructed earlier, we constructed a new dataset, which has more sequences. We achieved better prediction results using ID-SVM algorithm with our dataset, and obtained overall prediction accuracy of94.95%in Jackknife validation for the dataset of Du. The result was improved by5.3%and1.6%than AC and DWT algorithms respectively.2. On the bases of constructing the protein chemical shift dataset, we found that the four kinds of chemical shift of every amino acid relate with secondary structure and vary regularly after we had analyzed the relationship between secondary structure and the chemical shifts of backbone atoms for20amino acids. We achieved best results in predicting submitochondria location in present by constructing the protein feature vector using the auto covariance of chemical shifts.3. We proposed the amino acid stickiness feature vector, and achieved overall prediction accuracy of96.21%with Jackknife validation for the dataset of Du by using the stickiness, chemical shifts and other feature vectors. The result was improved1.26%than we studied before, and the protein located in matrix was predicted correctly. The prediction result of outer membrane was also improved by Sn of85.37%, which was about4.87%more than the best literature.4. We constructed the mycobacterial proteins dataset, and checked the generalization of extracting feature vector methods and prediction algorithm. The result was achieved94.00%by Jackknife validation and was2.8%and11.3higher than Lin’s and Rashid’s result respectively. The results show that our methods have strong generalization and can be used in other problems of subcellular locations.5. We established the chemical shifts algorithm-acACS web server (http://wlxy.imu.edu.cn/college/biostation/fuwu/acACS/index.asp) and submitochondria dataset web server (http://wlxy.imu.edu.cn/college/biostation/fuwu/mito/index.asp), which provide services for bioinformatics and protein submitochondria localization...