Predicting gene function using similarity learning

Phuong and Nhung BMC Genomics 2013, 14(Suppl 4):S4 http://www.biomedcentral.com/1471-2164/14/S4/S4 RESEARCH Open Access Predicting gene function using similarity learning Tu Minh Phuong1*, Ngo Phuong Nhung2 From IEEE International Conference on Bioinformatics and Biomedicine 2012 Philadelphia, PA, USA. 4-7 October 2012 Abstract Background: Computational methods that make use of heterogeneous biological datasets to predict gene function provide a cost-effective and rapid way for annotating genomes. A common framework shared by many such methods is to construct a combined functional association network from multiple networks representing different sources of data, and use this combined network as input to network-based or kernel-based learning algorithms. In these methods, a key factor contributing to the prediction accuracy is the network quality, which is the ability of the network to reflect the functional relatedness of gene pairs. To improve the network quality, a large effort has been spent on developing methods for network integration. These methods, however, produce networks, which then remain unchanged, and nearly no effort has been made to optimize the networks after their construction. Results: Here, we propose an alternative method to improve the network quality. The proposed method takes as input a combined network produced by an existing network integration algorithm, and reconstructs this network to better represent the co-functionality relationships between gene pairs. At the core of the method is a learning algorithm that can learn a measure of functional similarity between genes, which we then use to reconstruct the input network. In experiments with yeast and human, the proposed method produced improved networks and achieved more accurate results than two other leading gene function prediction approaches. Conclusions: The results show that it is possible to improve the accuracy of network-based gene function prediction methods by optimizing combined networks with appropriate similarity measures learned from data. The proposed learning procedure can handle noisy training data and scales well to large genomes. Background The increasing number of sequenced genomes makes it important to develop methods that can assign functions to newly discovered genes in a timely and cost-effective manner. Traditional laboratory methods, while accurate and reliable, would require enormous effort and time to identify functions for every gene. Computational approaches that utilize diverse biological datasets to generate automated predictions are useful in this situation as they can guide laboratory experiments and facilitate more rapid annotation of genomes. Existing computational approaches to gene function prediction have relied on a variety of genomic and * Correspondence: 1 Department of Computer Science, Posts & Telecommunications Institute of Technology, Hanoi, Viet Nam Full list of author information is available at the end of the article proteomic data. Exploiting the similarities between DNA or protein sequences to infer gene function was the first approach tested and has been the most widely used approach to date. Later, the usefulness of other types of genomic and proteomic data in this problem is also proved. Researchers have used microarray expression data [1], protein 3D structures [2], protein domain configuration [3], protein-protein interaction networks [4], and phylogenetic profiles [5] to predict functions of genes. Recently, inferring gene function simultaneously from different types of biological data has been shown to deliver more accurate predictions and has attracted considerable research interests [6-16]. Many methods for inferring functions of genes from heterogeneous datasets share a common framework in which a functional association between genes is first constructed and then used as input for learning algorithms. © 2013 Phuong and Nhung; licensee BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Phuong and Nhung BMC Genomics 2013, 14(Suppl 4):S4 http://www.biomedcentral.com/1471-2164/14/S4/S4 A functional association can be represented as a network with nodes corresponding to genes and edges representing the co-functionalities of gene pairs. In such a network, each edge is usually assigned a weight representing the strength of the co-functionality relationship between the gene pair. A network of this kind is typical constructed in two steps. First, each dataset is used to create an individual network that captures the co-functionality of gene pairs, as implied by this dataset. For vectorial data, one can calculate edge weights as the similarity scores between genes using appropriate similarity metrics, for example the Pearson correlation coefficient, and then form the networks by means of neighboring node connections. Data already given in forms of networks, for example protein-protein interactions, are used directly. The second step constructs a single combined association network by integrating the individual ones. A strategy commonly used in this step is to form the combined network as a weighted sum of individual ones. Here, each network is weighted according to its usefulness in predicting annotations for a group of genes that share a known specific function. Previous studies have used various regression or other learning based algorithms to estimate network weights. Given a functional association network, the next step is to use this network to propagate functional labels from a group of annotated genes to other genes. There are two main types of approaches for this step. Approaches of the first type create a kernel function from the co-functionality relationships encoded in the network and use this kernel with kernel-based classification algorithms [8,9,17]. In such approaches, genes with known annotations serve as labeled examples for training. Approaches of the second type use graph-based algorithms, which propagate labels from annotated genes to other genes based on graph proximity. Methods in this group range from simple nearest neighbor counting algorithms [16], to more sophisticated statistical methods such as graph-based semi-supervised learning algorithms [9], and Markov random fields [18] (see [19] for a more complete list of methods). On a number of benchmark datasets, graph-based and kernel-based approaches have shown comparable prediction accuracy, but graph-based approaches are generally faster [11,20]. The prediction accuracy of both graph-based and kernel-based approaches largely depends on the ability of the network to capture the functional associations between genes. To improve the network quality, previous studies have focused on improving the integration step, or more (...truncated)