Towards a Taxonomically Intelligent Phylogenetic Database
Roderic D. M. Page
Division of Environmental and Evolutionary Biology
Institute of Biomedical and Life Sciences
University of Glasgow
Glasgow, G12 8QQ, Scotland
r.page@bio.gla.ac.uk
Abstract
This note outlines some of the key intellectual obstacles that
stand in the way of creating a usable phylogenetic database.
These challenges include the need to accommodate multi-
ple taxonomic names and classifications, and the need for
tools to query trees in biologically meaningful ways. Until
these problems are addressed, and a taxonomically intelli-
gent phylogenetic database created, much of our phyloge-
netic knowledge will languish in the pages of journals.
1. Introduction
The last decade has seen an explosion in biological data, and
much of the success of the bioinformatics community has
derived from having ready access to that data. Databases
such as GenBank and EMBL have themselves spawned a
growing number of secondary databases tailored to specific
questions (such a protein and RNA structure, gene fami-
lies, metabolic pathways, polymorphisms, whole genomes,
etc.). Integrating these diverse databases has become a ma-
jor challenge for the bioinformatics community [18]. On-
tologies (controlled vocabularies) [2] and web services are
at the heart of projects such as BioMoby (biomoby.org)
and MyGrid (www.mygrid.org.uk), which aim to provide
descriptions of bioinformatics services and how to invoke
them [18]. The Science Environment for Ecological Knowl-
edge (SEEK) project has similar ambitions for ecological
data (EcoGrid) (seek.ecoinformatics.org).
The tree-building community has been noticeably ab-
sent from most of these developments. Although the phylo-
genetists have made considerable strides in the development
of sophisticated methods of phylogenetic inference, much
of the community's output of data sets and trees is languish-
ing in the pages of journals, rather than in openly accessi-
ble databases. Consequently, the rapid growth of published
evolutionary trees is not matched by the availability of those
Technical Reports in Taxonomy 04-01. Paper to be presented at the
Workshop on Database Issues in Biological Databases (DBiBD), National
e-Science Centre, Edinburgh, Scotland, January 8-9, 2005
1980
1985
1990
1995
2000
0
1000
2000
3000
4000
5000
6000
7000
Cumulative number
Year
Molecular phylogenies
TreeBASE studies
Figure 1: Cumulative growth of publications on phyloge-
netics, based on the number of papers found in the Web
of Science by searching on the key words "molecular" and
"phylogenetic" since 1981 [10], updated to 2003 and com-
pared with the growth of the phylogenetic repository Tree-
BASE, which launched in 1996 (a study in TreeBASE is
equivalent to a single paper).
trees in databases (Fig. 1). These trees and their support-
ing data form a tremendous resource for biologists, with ap-
plications in genomics, evolutionary biology, parasitology,
biodiversity, and public health [4]. The full potential of this
resource will only be realised when the phylogenetic com-
munity makes the results of their work more easily avail-
able.
Central to creating a useful phylogenetic database is hav-
ing ontologies for taxonomic names and characters, so that
meaningful queries can be constructed. Because trees are
complex structures (when compared, say, to the string of let-
ters representing a DNA sequence), we also need appropri-
ate ways to ask biologically interesting questions on trees.
This note explores some of these issues.
1
Nature Precedings : doi:10.1038/npre.2007.1028.1 : Posted 18 Sep 2007
2
The problem
There are several reasons why phylogenetic databases are
languishing behind sequence databases.
Most journals
make submission of sequences to public databases a pre-
requisite for publication, but very few impose a similar re-
quirement for trees and alignments (mycological journals
are a notable exception). There are also issues of computa-
tional infrastructure, staff, and long term funding that need
to be addressed. However, my goal here is to outline some
of the key intellectual obstacles that stand in the way of cre-
ating a usable database [6, 9, 8]. These challenges include:
(i) the lack of consistent taxonomic names; (ii) the absence
of standardised character names; (iii) the dearth of tools for
querying trees, and (iv) the lack of tools for synthesising
trees and datasets for large-scale analysis.
2.1
Where we are now: TreeBASE
The only viable phylogeny database currently available is
TreeBASE [12], which is housed at the University of Buf-
falo, New York. TreeBASE contains nearly 1000 studies
covering 39,000 taxa. In addition to a simple query in-
terface, TreeBASE provides some simple analytical tools,
including supertree construction using the modified min-
cut supertree algorithm [7] hosted on a machine at Glas-
gow. TreeBASE has also served as a test bed for a number
of theoretical developments in tree querying [15, 13, 20].
However, TreeBASE is greatly weakened by lacking a tax-
onomic framework.
2.2
Why taxonomy matters
Taxon names in a phylogenetic database should meet four
criteria: (i) internal consistency; (ii) external consistency;
(iii) synonymy, and (iv) hierarchy (10). The first criterion
(internal consistency) is an obvious requirement. If multiple
names are used for the same taxon, then a simple search for
all data relevant to a given taxon cannot be guaranteed to
have found all those data some might be associated with
an alternative name for that taxon.
The second criterion of external consistency assumes
that we want to be able to apply knowledge obtained from
the phylogenetic database to other domains. For example,
a user wanting to use phylogenetic methods to analyse the
evolutionary ecology of a group of organisms should be able
to use the same scientific name to obtain phylogenetic and
ecological data (for example through EcoGrid). This task
is complicated because the same taxon may have multiple
names (synonyms). A phylogeny database should be able
to translate between different names for the same taxon.
The final criterion of hierarchy is equivalent to requiring
an ontology that specifies the relationships between terms.
Figure 2: Graph of treespace in TreeBASE. Each node rep-
resents a study, and a pair of studies is connected by an
edge if the corresponding studies have at least one taxon
in common. Note that the graph is not connected. Studies
containing birds are indicated by
·.
For example, as text strings, "Gallus gallus" and "Struthio
camelus" have no obvious connection, but both are names
of birds (class Aves). If we query a phylogenetic database
using the term "Aves", we should be able to retrieve all stud-
ies containing a bird, regardless of whether those studies
actually contain a taxon labelled "Aves."
2.3
TreeBASE and taxonomy
TreeBASE has no mechanism for ensuring consistency of
names, nor does it have an ontology of taxonomic names.
These criteria are difficult to satisfy, and were especially so
at the time TreeBASE was designed (it went live in 1996).
TreeBASE tries to circumvent the need for a taxonomic hi-
erarchy using the notion of "tree surfing. The user is pre-
sented with a tree, and can "surf to neighbouring trees that
share at least one taxon in common with the original tree, in
the processes finding all the taxa of interest. However, for
tree surfing to work tree space in TreeBASE must be con-
nected, which it is not [11] (Fig. 2). Furthermore, even if it
were connected, there is no guarantee that all members of a
taxon will be contained in a set of neighbouring trees. In the
case of birds, the 24 studies in TreeBASE that contain one
or more birds lie in different components of the tree space
graph (Fig. 2), and so tree surfing by itself wont find all bird
phylogenies in TreeBASE.
3
Matching taxon names
3.1
Simple matching
A naive approach is to match names in TreeBASE to
names in an external database, such as the NCBI taxon-
2
Nature Precedings : doi:10.1038/npre.2007.1028.1 : Posted 18 Sep 2007
TreeBASE
TaxonName
Taxonomic
name database
(a)
(b)
(c)
2
1
3
4
Figure 3: Examples of matching names in TreeBASE to
TaxonNames in one or more taxonomic name databases. In
( a) the TaxonName matches a name in an external database;
in ( b) the TaxonName occurs in two different databases.
( c) shows a case where two different TaxonNames match
the same name in a taxonomic database (for example, the
TaxonNames comprise a taxon name concatenated with a
specimen number).
omy (www.ncbi.nlm.nih.gov) used in GenBank. Prelimi-
nary work shows that only around 45% of TreeBASE names
have an exact match to names in the NCBI taxonomy.
Because phylogenetic data come sources other than nu-
cleotide sequences, other taxonomic databases must be used
in addition to GenBank, such as ITIS (www.itis.usda.gov),
Species 2000 (www.sp2000.org), uBio (www.ubio.org),
IPNI (www.ipni.org), and Mammal Species of the World
(www.nmnh.si.edu/msw).
Some names in TreeBASE are spelt differently to names
in external databases, are informal names, or comprise sci-
entific names with voucher codes or GenBank accession
numbers tacked on the end. Hence, names will also have
to be matched using techniques such as approximate string
matching [21]. Where the name includes an accession num-
ber, this can be used to extract the organism name from Gen-
Bank.
We can model the problem of matching TreeBASE
names to taxonomic names using a bipartite graph, where
the nodes are partitioned into two disjoint sets, one repre-
senting names in TreeBASE, the other representing names
in taxonomic databases (Fig. 3). The components of the re-
sulting graph correspond to the sets of names that are equiv-
alent. For example, in this figure, the four TreeBASE Tax-
onNames belong in three components: 1, 2, and 3,4.
3.2
Synonyms
The mapping shown in Fig. 3 becomes more complicated
once we consider taxonomic synonyms and how they are
stored in different databases. For example, the harp seal
Phoca groenlandica
Pagophilus groenlandicus
Phoca groenlandica
180647
622022
T9589
Pagophilus groenlandicus
T9910
Pagophilus groenlandicus
Phoca groenlandica
39089
80725
Figure 4: Graph depicting the relationships between differ-
ent names for the harp seal. TreeBASE (
·) and GenBank ()
contain two different scientific names for this seal, and dif-
ferent data is associated with each name. However ITIS (
)
correctly links Phoca groenlandica and Pagophilus groen-
landicus as synonyms.
is known by two different scientific names, Phoca groen-
landica Erxlebben, 1777, and Pagophilus groenlandicus
(Erxleben, 1777).
These names are nomenclatural syn-
onyms: they refer to the same species, but place it in two
different genera. TreeBASE has both names and treats them
as different taxa, as does GenBank; however ITIS correctly
recognises that the two names are synonyms (Fig. 4).
Given cases like that in Fig. 4, we can no longer treat our
matching graph as bivariate (cf. Fig. 3), although the com-
ponents of that matching graph will still define sets of equiv-
alent names. However, the issue then arises "what name
to use for the harp seal? For internal use we can simply as-
sign a unique numerical identifier to each component, but if
the user wants to view information on a particular taxon (or
download that information into external software for further
analysis), they will need an "accepted name for each com-
ponent. We could represent the graph shown in Fig. 4 using
an ontology language such as RDF, and develop a set of
logical rules by which a reasoner (a program that explores
logical relationships) could compute an accepted name for
external use.
3.3
BLASTing TreeBASE
For those TreeBASE taxa with sequence data, an alternative
approach to matching names is to look up that sequence
in GenBank and compare the name of the corresponding
taxon in GenBank with that in TreeBASE. This involves
using BLAST [1] to find the best hit for each sequence in
TreeBASE. Care needs to be taken in the BLAST search
because some sequences in TreeBASE have not been de-
posited in GenBank, but will still return hits if GenBank
contains sequences from closely related taxa. Furthermore,
3
Nature Precedings : doi:10.1038/npre.2007.1028.1 : Posted 18 Sep 2007
Figure 5: Screenshot of the Glasgow Name Server (dar-
win.zoology.gla.ac.uk/ rpage/MyToL/www) displaying a
classification.
some TreeBASE datasets comprise concatenated genes, and
so only one of those genes will be represented in the top hit.
3.4
Taxonomy trees
Matching names is not enough to provide an intelligent tax-
onomic interface to TreeBASE. A user searching TreeBASE
for bird phylogenies using the term "Aves would find 4 stud-
ies, yet there are 24 bird studies in TreeBASE. Given that
tree surfing also fails to find all these studies (Fig. 2), we
need a different approach. Given a classification of all taxon
names in TreeBASE we would be able to compute the taxo-
nomic content of any name (for example, we would be able
to generate the list of all birds in TreeBASE).
A complication is that different taxonomic databases
employ different classifications. I have developed a pilot
database (the Glasgow Name Server) for storing, query-
ing and displaying classifications [8] using techniques de-
scribed by [3] (Fig. 5).
We need to explore different approaches to the problem
of combining multiple classifications, such as finding the
agreement between two classifications, or merging two or
more classifications into a consensus. These problems are
related to unordered subtree inclusion [19] and supertree
construction with internally labelled nodes [14], respec-
tively. Alternatively, perhaps multiple classifications could
be merged into a single graph that is no longer a tree.
4
Querying trees
There are two basic kinds of tree-based queries rele-
vant to phylogenetic databases:
pattern matching, and
finding trees that resemble another tree.
Tree pattern
matching techniques include ATreeGrep [13], and the
related TreeSearcher tool developed at Glasgow (dar-
win.zoology.gla.ac.uk/ rpage/TreeSearcher/).
To date all tree pattern matching queries have used ex-
isting labels in trees, that is, the queries make no use of
taxonomy. As a consequence, a simple query such as "find
all trees that have birds and crocodiles as sister taxa is dif-
ficult to formulate, unless the user knows the names of all
the birds and crocodilians in the database, and constructs
a suitable query tree (which may have hundreds of nodes).
To avoid this problem, we need to develop a query rewrit-
ing mechanism that uses a taxonomic classification and a
mapping of tree labels to taxon names o rewrite the query.
Hence, the user could enter a query such as find phyloge-
nies matching the tree ((Aves,,Crocodilia),Testudines), and
get a meaningful answer, even if each tree in the database is
labelled only with species names.
4.1
Finding trees that resemble a query tree
Given a tree, such as one in TreeBASE, a natural question
to ask is "are there any other trees that look like this? To an-
swer this, we need a tree comparison measure, and a means
of quickly comparing our tree with the other trees in Tree-
BASE. There are numerous tree comparison metrics avail-
able [17], some of which have been developed with search-
ing tree databases in mind [20].
The simple approach to finding similar trees is to do a
linear search, that is, compare the query tree with every tree
in the database. In practice this approach may perform rea-
sonably well, especially if the database is small and we can
filter out lots of trees (for example, those that have no taxa
in common with the query tree). However, as the database
grows this method is unlikely to scale.
Another approach is to define a metric on tree space,
and use a metric-space index to quickly find similar trees.
Metric-space indices are a very attractive means of search-
ing complex data types, such as sequences, images, and
trees [5]. Although there is an extensive literature on tree
comparison measures, there are no published metrics for
trees where the trees have different, possibly overlapping
leaves. However, we can use a symmetric difference as a
starting point. Given two trees, T
1
and T
2
, then
d
(T
1
, T
2
) = |T
1
| + |T
2
| - 2|T
1
T
2
|
where
|T
1
| is the number of leaves in the tree T
1
. We can
normalise this by dividing it by
|T
1
| + |T 2|, so that it ranges
from 0 to 1. This measure regards trees as similar if they
have taxa in common, but makes no use of tree topology.
We can incorporate tree topology using maximum agree-
ment substrees (MAST) [16]
d
(T
1
, T
2
) = |T
1
| + |T
2
| - 2 × MAST(T
1
, T
2
)
4
Nature Precedings : doi:10.1038/npre.2007.1028.1 : Posted 18 Sep 2007
where
MAST(T
1
, T
2
) is the number of taxa in the max-
imum agreement subtree.
Implementations of metric indices exist (e.g., the C++
library M-tree library: www-db.deis.unibo.it/Mtree/), and
it would be interesting to evaluate the performance of this
approach to searching for trees.
References
[1] S.F. Altschul, W. Gish, W. Miller, E. W. Myers, and
D. J. Lipman. Basic local alignment search tool. Jour-
nal of Molecular Biology, 215:403410, 1990.
[2] J.B.L. Bard and S.Y. Ree. Ontologies in biology: de-
sign, applications, and future challenges. Nature Re-
views Genetics, 5:213222, 2004.
[3] J. Celko. Joe Celko's Trees and Hierarchies in SQL for
Smarties. Morgan Kaufmann, San Francisco, 2004.
[4] J. Cracraft, M.J. Donoghue, J. Dragoo, D. Hillis, and
T. Yates. Assembling the Tree of Life: Harnessing
Life's History to Benefit Science and Society. Ameri-
can Museum of Natural History, 2003.
[5] D. Miranker. Metric-space indexes as a basis for scal-
able biological databases. Omics: A Journal of Inte-
grative Biology, 7:5760, 2003.
[6] L. Nakhleh, D. Miranker, F. Barbancon, W.H. Piel,
and M.J. Donoghue. Requirements of phylogenetic
databases. In Third IEEE Symposium on BioInformat-
ics and BioEngineering (BIBE'03), pages 141148,
2003.
[7] R.D.M. Page. Modified mincut supertrees. In Pro-
ceedings of the Second Workshop on Algorithms in
Bioinformatics (WABI2002) (Lecture Notes in Com-
puter Science 2452), Rome, Italy, 2002.
[8] R.D.M. Page. Phyloinformatics: Towards a phyloge-
netic database. In J. T. L. Wang, M.J. Zaki, H.T.T.
Toivonen, and D. Shasha, editors, Data Mining in
Bioinformatics Data Mining in Bioinformatics, Ad-
vanced Information and Knowledge Processing, pages
219241. Springer-Verlag, 2004.
[9] R.D.M. Page. Taxonomy, supertrees, and the tree of
life. In O. Bininda-Emonds, editor, Phylogenetic su-
pertrees: Combining information to reveal the Tree
of Life, volume 4 of Computational Biology. Kluwer
Academic Publishers, 2004.
[10] M. Pagel. Inferring the historical patterns of biological
evolution. Nature, 401:877884, 1999.
[11] W. Piel, M.J. Sanderson, and M.J. Donoghue. The
small-world dynamics of tree networks and data min-
ing in phyloinformatics.
Bioinformatics, 19:1162
1168, 2003.
[12] W.H. Piel, M.J. Donoghue, and M.J. Sanderson. Tree-
BASE: a database of phylogenetic knowledge.
In
To the Interoperable Catalogue of Life with part-
ners Species 2000 Asia Oceania Proceedings
of 2nd International Workshop of Species 2000 (Re-
search Report for the National Institute of Envi-
ronmental Studies, R-171-2002), Tsukuba, Japan,
2002. http://www.nies.go.jp/kanko/kenkyu/pdf/r-171-
2002.pdf.
[13] D. Sasha, J.T.L. Wang, H. Shan, and K. Zhang. Atree-
grep: approximate searching in unordered trees. In
J. Kennedy, editor, Scientific and Statistical Database
Management (SSDBM 2002), pages 8998, 2002.
[14] C. Semple, P. Daniel, W. Hordijk, R.D.M. Page, and
M. Steel.
Supertree algorithms for ancestral diver-
gence dates and nested taxa. Bioinformatics, in press.
[15] H. Shan, K.G. Herbert, W. H. Piel, D. Sasha, and
J.T.L. Wang. A structure-based search engine for phy-
logenetic databases.
In 14th International Confer-
ence on Scientific and Statistical Database Manage-
ment (SSDBM'02), pages 710, 2002.
[16] M. Steel and T. Warnow.
Kaikoura tree theorems:
computing the maximum agreement subtree. Infor-
mation Processing Letters, 48:7782, 1993.
[17] M.A. Steel and D. Penny. Distributions of tree com-
parison metrics some new results. Systematic Biol-
ogy, 42:126141, 1993.
[18] Lincoln Stein. Integrating biological databases. Na-
ture Reviews Genetics, 4:337345, 2003.
[19] G. Valiente. Algorithms on trees and graphs. Springer-
Verlag, Heidelberg, 2002.
[20] J.T.L. Wang, H. Shan, and D. Sasha abd W.H. Piel.
Treerank: a similarity measure for nearest neighbor
searching in phylogenetic databases.
In Proceed-
ings of the 15th International Conference on Scien-
tific and Statistical Database Management (SSDBM
2003), pages 171180, 2003.
[21] S. Wu and U. Manber. Agrep a fast approximate
pattern-matching tool. In Proceedings USENIX Win-
ter 1992 Technical Conference, pages 153162, San
Francisco, CA, 1992.
5
Nature Precedings : doi:10.1038/npre.2007.1028.1 : Posted 18 Sep 2007