Infectious pathogens such as influenza virus and HIV have large databases of genetic data. In spite of efforts to sequence broadly representative viruses, bias still emerges from the substantial contribution of data from specialized sources, such as infection outbreaks, vaccine candidate selection or clinical studies. When databases are large and imbalanced, they pose a substantial challenge to selection of a dataset with desired evolutionary properties, such as evolutionary representation or shared genotype. An obvious selection approach is to control for relevant epidemiological variables. However, the resulting dataset may not have the desired properties, particularly when organisms are readily transmitted. For example, two identical (or almost identical) human influenza A viruses can be collected from hosts separated by thousands of miles. An alternative selection approach is to use an alignment of candidate nucleotide or protein sequences for visualizing genetic relationships. However, alignments are viewed as deviations from a reference sequence, making it easy to see when a sequence is similar to the reference sequence, but difficult to visualize the degree of similarity among non-reference sequences.

When data sets are imbalanced, clustering into groups of similar sequences is a powerful adjunct to the selection process. An example of such a capability for influenza sequences is SmartBLAST [1]. SmartBLAST uses a query sequence to search the database for other sequences within a close distance, as measured by BLAST. SmartBLAST then collapses the ordered BLAST hits into clusters of similar sequences and chooses a representative of each cluster to comprise a dataset for subsequent analysis. Currently, SmartBLAST operates on protein sequences and the degree of similarity that determines cluster membership is preset. Performance is achieved by pre-computing pairwise BLAST similarity scores between all pairs of sequences in the influenza database.

SmartBLAST is well suited for selecting a coarse representation from a very large data set. For example, SmartBLAST with a recent (2009) influenza A hemagglutinin (H1) query sequence efficiently selects 120 representatives from 3000+ hemagglutinin (H1) sequences from viruses collected between 1918 and the present. However, there are tradeoffs for the simplicity and efficiency of the automation of SmartBLAST. For example, SmartBLAST does not allow the user to control the degree of similarity required for clustering. In the above example of SmartBLAST, all 2009 pandemic H1 sequences fell into a single cluster. Further, SmartBLAST operates on protein sequences. Therefore, similarity in SmartBLAST does not guarantee closeness in a nucleotide-based phylogenetic tree. If a cluster can be equally well represented by multiple sequences, the user cannot choose from alternatives based on relevant epidemiological or other variables. Automation also denies the user the possibility of recognizing and culling data with incorrect details. SmartBLAST does not allow the user to define the domain within which selections are to be made. For example, a user may be interested in only sequences from a specific host species. Such epidemiological control is not possible in SmartBLAST. For the above reasons, we believe that SmartBLAST generates data sets well suited to exploratory data analysis, but possibly not well suited for detailed analysis of specific evolutionary events.

To select a dataset with the desired evolutionary properties, selection from a phylogeny of the candidate data provides a direct, and therefore efficient, approach. The adaptive tree visualization capability [2] of the Influenza Virus Resource [3] can vary the amount of detail presented in a phylogenetic tree display. Tree-distance between sequences is used to automatically collapse sub-trees to a degree of detail controlled by the user. However, the visualization does not provide a mechanism for producing a dataset to match the tree as ultimately visualized. As with SmartBLAST, the automation of the adaptive tree visualization has tradeoffs for the user, who gives up control over the selection of the representatives of a sub-tree.

We developed a tool, Tree Pruner, for obtaining a dataset with desired evolutionary properties from a large, biased genetic database. Tree Pruner produces a high-quality dataset by giving the user complete control over the data selection process. Thus, Tree Pruner trades off some of the simplicity and speed of automation in the interest of greater quality of the final data selection. Data selection is carried out by interaction with a display of the phylogeny of a large selection of data from a domain of interest. Custom Tree Pruner commands allow the user to fade-out or accentuate tips or sub-trees in the tree display to represent de-selection or selection of data to be included in the final dataset. Thus, at all times the user has a view of the approximate tree of selected sequences only. Because Tree Pruner gives the user complete control over the choice of evolutionarily representative sequences, it fosters selection of high-quality datasets. When Tree Pruner is connected to data management capabilities, such as those of the Influenza Research Database, the user can save an edited dataset for further editing or subsequent analysis.

Tree Pruner is an efficient, visual editing capability for obtaining a dataset of genetic sequences with desired properties, such as evolutionary representation or shared genotype. Importantly, it provides the user curatorial control over the final selection of sequence data. While it is currently used with the large, biased influenza sequence database, it can be implemented for other viral genetic databases, such as those for HIV and HCV.

Example of editing using Tree Pruner

The following illustrations of Tree Pruner are based on an initial dataset of all instances of sequences of the hemagglutinin (HA) gene from influenza A (H5N1) viruses that were collected in the period 1900 - 2000. A minimum length of 900 (out of a maximum of 1790) nucleotides was required for inclusion in this initial dataset. A search of the Influenza Research Database (IRD) [4, 5], conducted on March 7 2010, yielded 96 sequences, which were stored in a working set called DemoTreePruner on the IRD server. Instructions for accessing this dataset to test Tree Pruner are given in Figure 1. (Alternatively, a static version of Tree Pruner with a pre-loaded tree of 953 NP sequences from seasonal influenza A (H1N1) viruses, together with an equine NP sequence as an outgroup, is available at the Influenza Sequence Database, http://www.flu.lanl.gov.)

Tree Pruner is launched from the working set, and automatically infers a phylogeny of the sequences in the set. (The IRD infers a maximum likelihood tree under the HKY model of evolution, using PhyML [10, 11].) Tree Pruner then opens an Archaeopteryx applet, labeled with the name of the working set (DemoTreePruner) and the name of the gene (Segment 4), and displays this tree.

The user can carry out multiple Keep and/or Remove editing sessions. Before switching from one edit action to the other, the user must save or discard all edits of the current type. This ensures that the keep/remove status of all "untouched" taxa is defined. The user can end a Tree Pruner session by committing changes. Then taxa marked for removal in the tree will be removed from the dataset, thus maintaining a 1:1 relationship between the dataset and tree display.

(i) Keep function

When a user clicks on a node at the tip of the tree, or at the root of a sub-tree, the label(s) of the selected tip (sub-tree) is (are) written in black, designating inclusion in the final dataset. All non-selected labels are written in blue. Blue tips are "active," meaning they can be selected by future actions, but are not currently designated for inclusion in the final dataset. Branches are also colored; black branches lead to sequences selected for inclusion in the dataset; blue branches lead to sequences whose status is undetermined. (See Figure 2.)

(ii) Remove/Restore function

When a user clicks on a node at the tip of the tree, or at the root of a sub-tree, the color of the label(s) of the selected tip (sub-tree) switches between black and grey. All non-selected labels remain black. Black tips are "active" and can be removed by future actions. Alternatively, clicking on a grey node is a Restore action. Restore will change the tip label(s) and relevant branches from grey to black. (See Figure 3.)

(iii) Handling large trees

A key scientific requirement of Tree Pruner was to enable the user to view taxon labels, even in large trees. Archaeopteryx offers three features for viewing large trees. "Dynamic hiding" displays a subset only of taxon labels in order to squeeze a large tree into the window. Zooming allows reading of all taxon labels, but loses the overall tree structure. Viewing a sub-tree allows reading of all taxon labels in a sub-tree, but loses the context of editing because the sub-tree replaces the full tree in the window. Archaeopteryx was modified so that Tree Pruner displays a sub-tree in a separate window, side-by-side with the display of the complete tree. Multiple levels of sub-tree are permitted. Pruning actions are mapped among windows by the Refresh action. Thus, a sub-tree can be pruned while viewing its context.

(iv) Concluding a Tree Pruner session

The user may use Commit to exit from Tree Pruner; all taxa tagged for removal (i.e., drawn in grey in the phylogeny) will be removed from the working set. The tree used to edit the dataset is removed from the server. To perform further editing, Tree Pruner infers a tree from the revised dataset, thereby retaining a 1:1 correspondence between the dataset and the tree display.

(v) Crash recovery

AutoSave is run every 10 minutes after the most recent user-instigated Save. If the server or client crashes during a pruning session, Tree Pruner can resume from the most recent (Auto)Save.

Selecting a dataset of sequences with specific evolutionary properties from a large, biased genetic database is problematic. Selection by inspection and editing of a phylogeny of candidate sequences is an obvious approach. Tools such as TreeDyn [7] and Archaeopteryx [6] allow manual editing of a phylogenetic display; the Influenza Virus Resource's [3] adaptive tree visualization [2] uses an algorithm to automatically edit a phylogenetic display. However, none of these tools can translate editing actions into data selection.

Tree Pruner was designed for selecting data from the influenza genetic sequence database. Tree Pruner uses a manual editing process to avoid the potential of automatic editing to select poor quality data. The intuitive editing actions of Tree Pruner allow the user to quickly converge on a final data set with the required evolutionary properties. Tree Pruner uses sub-trees to facilitate reading the labels on a large tree, thus fostering curation of the final data. Communications between the Tree Pruner applet and the server translate editing of the phylogeny into selection of data from the database.

The size of the dataset that can be edited by Tree Pruner is limited by the capacity of the server to infer a phylogeny for the initial superset of data and not by the capacity of Archaeopteryx, which can display trees of 5,000 - 10,000 taxa without significant degradation of performance. Currently, PhyML is used to infer trees for Tree Pruner; PhyML works reliably for trees with 1,000 - 2,000 taxa. Thus, Tree Pruner is relevant for typical large influenza virus sequence analyses.

Tree Pruner can be easily adapted for data selection from large genetic databases of other organisms such as HIV, provided the appropriate software for applet-server communications is used, and the database has a means of storing a dataset.

Tree Pruner is an efficient tool for selecting a curated dataset with the required evolutionary properties from a large, biased database of genetic sequences.

Project name: Tree Pruner

Code availability: http://github.com/caml/Tree-Pruner/

Link to full implementation: http://www.fludb.org/

Link to static demonstration: http://www.flu.lanl.gov

Link to user guide: http://www.fludb.org/brcDocs/tutorials/TreeEditor.pdf

http://www.flu.lanl.gov/

Operating systems Platform-independent if browser is Java-compatible

Programming language: Java

Other requirements: A code-signing certificate is required to allow exporting of graphics onto a user's desktop.

License: L-GPL