PMC:539961 JSONTXT 3 Projects

ProtoNet 4.0: A hierarchical classification of one million protein sequences Abstract ProtoNet is an automatic hierarchical classification of the protein sequence space. In 2004, the ProtoNet (version 4.0) presents the analysis of over one million proteins merged from SwissProt and TrEMBL databases. In addition to rich visualization and analysis tools to navigate the clustering hierarchy, we incorporated several improvements that allow a simplified view of the scaffold of the proteins. An unsupervised, biologically valid method that was developed resulted in a condensation of the ProtoNet hierarchy to only 12% of the clusters. A large portion of these clusters was automatically assigned high confidence biological names according to their correspondence with functional annotations. ProtoNet is available at: http://www.protonet.cs.huji.ac.il. INTRODUCTION ProtoNet (1) (launched in 2002) is an automatic hierarchical clustering of the SwissProt and TrEMBL (2) protein databases. The clustering process is based on an all-against-all BLAST (3) similarity search. The similarities' E-score is used to perform a continuous bottom-up clustering process by joining the two most similar protein clusters at each step, resulting in a hierarchy of protein clusters at various degrees of biological granularity. This hierarchy is structured as a collection of trees, in which the root clusters contain all the proteins of the tree and the rest of the clusters represent subdivisions of the proteins into smaller groups. Several interesting biological insights about protein families and the evolution of structural and functional relations between proteins are obtained by browsing this global hierarchical organization of the protein world (4). Furthermore, ProtoNet can be used to assess the function of novel protein sequences, by finding the best matching cluster for the new sequence. In this paper, we describe several new developments in ProtoNet 4.0 including the increase of the sequences analyzed from 114 033 proteins in SwissProt (version 2.1) to 1 072 911 sequences (SwissProt and TrEMBL, version 4.0) and the improved methodology for simplification of the scaffold of the protein hierarchy. ProtoNet is available at: http://www.protonet.cs.huji.ac.il. HIERARCHY CONDENSATION Due to the immense size of the ProtoNet hierarchy and the number of protein clusters, it would be very difficult to navigate in such a large hierarchy. Furthermore, it is obvious that many of the clusters are biologically irrelevant and uninteresting (e.g. huge root clusters that contain hundreds of thousands non-related proteins). In order to get a condensed yet biologically relevant view of the hierarchy, we searched for some process-intrinsic parameter that would indicate the type of clusters that are biologically relevant. The parameter found measures the stability of the cluster in the process, assuming that a stable cluster would be also more relevant biologically. We found this assumption to be correct, and that if we select a small subset of the clusters that show high stability, they would retain the biological validity of ProtoNet (N. Kaplan et al., manuscript in preparation). The default condensation of the ProtoNet hierarchy leaves 12% of the clusters. However, the ProtoNet website now offers an ‘advanced mode’, in which advanced users can control the level of condensation and the method by which it is done, which results in a larger or smaller set of clusters as required. It is to be noted that the condensation causes the trees to change from binary trees to non-binary trees, and some browsing options have been developed accordingly (see ‘Web Enhancements’). DATABASE UPDATES ProtoNet has gone through a major update of all database sources. Primarily, the protein database from which the ProtoNet tree is constructed has been updated and extended to include the TrEMBL protein database as well as SwissProt. This results in a leap from 114 033 to 1 072 911 protein sequences. Although TrEMBL is not manually validated by experts, it provides a much more extensive view of the protein world including whole genomes and thorough representation of several key organisms (see Table 1). CLUSTER NAMES A major objective in bioinformatics is to assign a biological function to proteins. We have developed an automatic high-confidence method that assigns a functional annotation to ProtoNet protein clusters. The method finds a functional annotation from either InterPro (5), GO (6), SwissProt or ENZYME (7) databases that best fits the proteins of the cluster and assigns it a score relative to how well it fits the cluster. If this score is above a certain threshold, the annotation is assigned as the cluster's name. Understandably, not all protein clusters would have an existing annotation that fits them well. Clusters whose best fitting annotation does not pass the threshold remain nameless, which possibly suggests a novel functional group or clusters that are associated with mixed functions. By applying this method, we were able to assign biological names to 78% of the clusters that contain 20 proteins or more. The annotation can be assigned with high confidence to the cluster because a high threshold is used. WEBSITE ENHANCEMENTS Several enhancements have been made to the ProtoNet website in order to allow easier and more in-depth analysis of the ProtoNet trees. Browsing cluster names Cluster names are extremely useful for quickly browsing the ProtoNet trees, which eliminates the need to check the proteins of each cluster in order to get an impression of the hierarchy. Furthermore, the assignment of a biological function to clusters suggests an easy scheme of assigning a function to proteins with unknown function: a protein can be assigned the function of all clusters to which it belongs. This scheme can be used not only on each of the 1 072 911 proteins from which the ProtoNet hierarchy is constructed, but also for new protein sequences given by the user (using the ‘Classify your protein’ option in the website, which finds the most suitable cluster for a new protein sequence given by the user). Browsing the tree Subtree view In order to cope with the change to a non-binary tree, we have introduced the ProtoBrowser (Figure 1), which shows the tree in the vicinity of the cluster that is being displayed. Instead of presenting only the branch of the tree to which the cluster belonged to, the new display allows an easy navigation to neighboring clusters and an enhanced global overview of biological protein families. Functionality view PANDORA (8) is a web-based tool that allows an in-depth biological analysis of large protein sets. It is a natural choice when trying to biologically interpret large protein clusters that contain hundreds of proteins. ProtoNet now offers a direct link from its cluster page to PANDORA, providing the ability to easily understand what biological groups the cluster is built from and analyze them from several different biological aspects. Similarity view When viewing a protein cluster, it is sometimes helpful to obtain an in-depth look into the sequence similarity between the proteins of the cluster. This could allow the user to identify if there is a natural partitioning into subgroups or if the inner similarity of the cluster is uniform. In order to address this, the website offers a cluster similarity matrix (Figure 2), which shows an all-against-all color-coded matrix of all protein pairs in the cluster, colored according to the BLAST E-score between the two proteins. This also facilitates access to the BLAST result, which can be obtained simply by clicking on the appropriate cell in the matrix. Maintenance and updating The ProtoNet source databases are generally updated twice a year. The next ProtoNet release is planned to include the UniProt Ref 100 protein database. Other future plans include allowing the users to select a subset of proteins from ProtoNet according to their needs (e.g. selecting for study only the proteins from the SwissProt database or only the proteins of a specific species) and expanding links to further biological databases such as OMIM (9) and DIP (10).