Munira Alballa1,2, Gregory Butler3,4. 1. Department of Computer Science and Software Engineering, Concordia University, Montreal, QC, Canada. m_alball@encs.concordia.ca. 2. College of Computer and Information Sciences, King Saud University, Riyadh, Saudi Arabia. m_alball@encs.concordia.ca. 3. Department of Computer Science and Software Engineering, Concordia University, Montreal, QC, Canada. 4. Centre for Structural and Functional Genomics, Concordia University, Montreal, QC, 24105, Canada.
Abstract
BACKGROUND: Membrane proteins are key gates that control various vital cellular functions. Membrane proteins are often detected using transmembrane topology prediction tools. While transmembrane topology prediction tools can detect integral membrane proteins, they do not address surface-bound proteins. In this study, we focused on finding the best techniques for distinguishing all types of membrane proteins. RESULTS: This research first demonstrates the shortcomings of merely using transmembrane topology prediction tools to detect all types of membrane proteins. Then, the performance of various feature extraction techniques in combination with different machine learning algorithms was explored. The experimental results obtained by cross-validation and independent testing suggest that applying an integrative approach that combines the results of transmembrane topology prediction and position-specific scoring matrix (Pse-PSSM) optimized evidence-theoretic k nearest neighbor (OET-KNN) predictors yields the best performance. CONCLUSION: The integrative approach outperforms the state-of-the-art methods in terms of accuracy and MCC, where the accuracy reached a 92.51% in independent testing, compared to the 89.53% and 79.42% accuracies achieved by the state-of-the-art methods.
BACKGROUND: Membrane proteins are key gates that control various vital cellular functions. Membrane proteins are often detected using transmembrane topology prediction tools. While transmembrane topology prediction tools can detect integral membrane proteins, they do not address surface-bound proteins. In this study, we focused on finding the best techniques for distinguishing all types of membrane proteins. RESULTS: This research first demonstrates the shortcomings of merely using transmembrane topology prediction tools to detect all types of membrane proteins. Then, the performance of various feature extraction techniques in combination with different machine learning algorithms was explored. The experimental results obtained by cross-validation and independent testing suggest that applying an integrative approach that combines the results of transmembrane topology prediction and position-specific scoring matrix (Pse-PSSM) optimized evidence-theoretic k nearest neighbor (OET-KNN) predictors yields the best performance. CONCLUSION: The integrative approach outperforms the state-of-the-art methods in terms of accuracy and MCC, where the accuracy reached a 92.51% in independent testing, compared to the 89.53% and 79.42% accuracies achieved by the state-of-the-art methods.
Authors: Andreas Bernsel; Håkan Viklund; Jenny Falk; Erik Lindahl; Gunnar von Heijne; Arne Elofsson Journal: Proc Natl Acad Sci U S A Date: 2008-05-13 Impact factor: 11.205