Classification of protein folds plays a very important role in the protein structure discovery process, especially when traditional sequence alignment methods fail to yield convincing structural homologies. In this chapter, we have developed a two-layer learning architecture, named TLLA, for multi-class protein folds classification. In the first layer, OET-KNN (Optimized Evidence-Theoretic K Nearest Neighbors) is used as the component classifier to find the most probable K-folds of the query protein. In the second layer, we use support vector machine (SVM) to build the multi-class classifier just on the K-folds, generated in the first layer, rather than on all the 27 folds. For multi-feature combination, ensemble strategy based on voting is selected to give the final classification result. The standard percentage accuracy of our method at ~63% is achieved on the independent testing dataset, where most of the proteins have <25% sequence identity with those in the training dataset. The experimental evaluation based on a widely used benchmark dataset has shown that our approach outperforms the competing methods, implying our approach might become a useful vehicle in the literature.
TopIntroduction
The successful completion of many genome sequencing projects produces a massive number of putative protein sequences. However, the number of known three-dimensional (3D) protein structure is growing at a much slower pace. This situation has challenged us to develop computational methods by which the 3D protein structure could be predicted timely from its sequence.
Many computational methods have been developed, which are used for assigning folds to protein sequences. Traditional similarity-based methods, such as sequence-structure homology recognition methods (Shi et al., 2001), align target sequence onto known structural templates and calculate their sequence-structure compatibilities using either profile-based scoring functions (Kelley et al., 2000) or environment-specific substitution tables (Shi et al., 2001). The scores obtained for different structural templates are then ranked. And the template, which gives rise to the best score, is assumed to be the fold of the target sequence. Although these methods can accurately recognize protein folds when proteins have close evolutionary relationship (Yu et al., 2006), they are not efficient when proteins are structurally similar, but have no significant sequence similarity. Threading methods have demonstrated promising results in detecting the latter type of relationship. However, these methods also yield poor accuracies perhaps due to the difficulty of formulating reliable and general scoring functions (Jones, 1999).
In recent years, the taxonomic approach without relying on sequence similarity, such as those using machine learning methods (Ding & Dubchak, 2001; Huang et al., 2003; Nanni, 2006; Shen & Chou, 2006; Guo & Gao, 2008), plays a critical role in protein folds recognition. This approach assumes that the number of protein folds is limited according to the SCOP (Structural Classification of Proteins) database (Andreeva et al., 2004), therefore, the protein fold recognition problem can be viewed as a fold classification problem, where a query protein can be classified into one of the known folds. Ding and Dubchak (2001) investigated support vector machine (SVM) and artificial neural network (ANN) methods for protein folds classification. Shen and Nanni studied ensemble classifiers based on nearest neighbor and K-local hyperplane (Shen & Chou, 2006; Nanni, 2006), respectively. Guo and Gao (2008) developed a novel ensemble classifier using GAET-KNN. These taxonomic approaches have achieved much better performance than the traditional similarity-based methods.
Most current taxonomic approaches assign the fold of a query protein to one of the most populated 27 folds in the SCOP (such as globin-like, cupredoxins, beta-trefoil, etc.). Protein folds prediction in the context of this large number of folds presents a rather challenging classification problem. Huang and Lin (2003) developed a hierarchical learning architecture (HLA). In the first level of HLA, it classifies the protein sequence into four major structural classes. And in the next level, it further classifies the protein sequence into 27 folds. This two step strategy reasonably reduced the number of folds to classify. HLA obtained higher accuracy than those methods which classify the query protein into 27 folds in one step. Hence, reducing the number of probable folds of the query protein is instructive to achieve better accuracy. Toward this goal, in the article, a two-layer learning architecture (TLLA) is presented. In the TLLA, we first use Optimized Evidence-Theoretic K Nearest Neighbors (OET-KNN) to find the most probable K-folds, then build the multi-class classifier just on the K-folds based on SVM. Finally, the query protein is tested on the multi-class classifier and its fold type is predicted. The standard percentage accuracy of our method at ~63% is achieved on a widely used independent testing dataset, which is better than the competing methods. Therefore, we suggest that TLLA can be used for fold-wise classification of unknown proteins discovered in various genomes.