A robust method to detect structural and functional remote homologues.

Abstract

With currently available sequence data, it is feasible to conduct extensive comparisons among large sets of protein sequences. It is still a much more challenging task to partition the protein space into structurally and functionally related families solely based on sequence comparisons. The ProtoNet system automatically generates a treelike classification of the whole protein space. It stands to reason that this classification reflects evolutionary relationships, both close and remote. In this article, we examine this hypothesis. We present a semiautomatic procedure that singles out certain inner nodes in the ProtoNet tree that should ideally correspond to structurally and functionally defined protein families. We compare the performance of this method against several expert systems. Some of the competing methods incorporate additional extraneous information on protein structure or on enzymatic activities. The ProtoNet-based method performs at least as well as any of the methods with which it was compared. This article illustrates the ProtoNet-based method on several evolutionarily diverse families. Using this new method, an evolutionary divergence scheme can be proposed for a large number of structural and functional related superfamilies.

Full text

Page 1

A Robust Method to Detect Structural and Functional
Remote Homologues
Ori Shachar1and Michal Linial2*
1School of Computer Science and Engineering, Hebrew University, Jerusalem, Israel
2Department of Biological Chemistry, Institute of Life Sciences, Hebrew University, Jerusalem, Israel
ABSTRACT
data, it is feasible to conduct extensive comparisons
among large sets of protein sequences. It is still a
much more challenging task to partition the protein
spaceintostructurallyandfunctionallyrelatedfami-
lies solely based on sequence comparisons. The
ProtoNet system automatically generates a treelike
classification of the whole protein space. It stands to
reason that this classification reflects evolutionary
relationships, both close and remote. In this article,
we examine this hypothesis. We present a semiauto-
maticprocedurethatsinglesoutcertaininnernodes
in the ProtoNet tree that should ideally correspond
tostructurallyandfunctionallydefinedproteinfami-
lies. We compare the performance of this method
against several expert systems. Some of the compet-
ing methods incorporate additional extraneous in-
formation on protein structure or on enzymatic
activities. The ProtoNet-based method performs at
least as well as any of the methods with which it was
compared. This article illustrates the ProtoNet-
basedmethodonseveralevolutionarilydiversefami-
lies. Using this new method, an evolutionary diver-
gence scheme can be proposed for a large number of
structural and functional related superfamilies.
Proteins 2004;57:531–538.
With currently available sequence
© 2004 Wiley-Liss, Inc.
Key words: protein family; hierarchical classifica-
tion; protein space; database
INTRODUCTION
The gap between the amount of sequences available and
the number of known homologous families is still grow-
ing.1,2This article addresses the problem of finding remote
homologues in the sequence space and suggesting for any
given protein sequence its relationship with other pro-
teins, which basically form a structural or a functional
homologous family.
To fully appreciate the complexity of the problem, one
needs to imagine the whole sequence space of biologically
functional sequences, where the pairwise distance be-
tween every pair of sequences is exactly the pairwise score
given by some chosen method for sequence similarity (i.e.,
FASTA or BLAST, with any chosen substitution matrix3).
Our task of defining functional families would sum up to
drawing an outline around all proteins that belong to the
same functional family, thus separating them from all
sequences of other families. For single-domain proteins
that have evolved by divergent evolution, such a descrip-
tion is valid, as in the case of the globin and histone
families. However, for multidomain proteins that may
have resulted from the fusion of domains, and for those
that have evolved by convergent evolution,4such a simpli-
fied partition of the protein sequence space is incompat-
ible. Such groups of proteins that are related by conver-
gence in evolution or that are linked through shared
domains are abundant in the protein space. They are
especially abundant in complex proteomes. In reality,
methods for classifications of such cases into functional or
structural families are not reliable.
Most proteins, especially those derived from newly
sequenced genomes, are not yet annotated, and their
evolutionary relations to other proteins are not evident. As
a result, the task of expanding and generalizing the
presently available classification schemes for new se-
quences is difficult and may result in a reduction of the
quality of functional inference.
The basic principle for classifying a new sequence is to
test its identity and similarity to all other sequences in the
database. Should it have a sequence identity of at least
30%withanothersequence(throughoutmostofitslength),
a structural or a functional inference becomes valid. More
accurately, functional annotation such as catalytic func-
tion of an enzyme requires at least 40% identity in amino
acids,5while reliable structural inference can be based on
25–30% sequence identity over at least 100 amino acids.6
Below these levels of pairwise sequence similarity, no
unsupervised search method can suggest, with a reason-
able confidence, the structural or functional relatedness to
the closest homologue. While methods based on BLAST-
basediterativeprofiles,HiddenMarkovModels,andmeth-
ods for finding connections through intermediate se-
quences are superior in their sensitivity, they suffer from a
higher level of false positives.1Specifically, PSI-BLAST,
which is among the best search methods,7misses ?50% of
Grant sponsor: Israeli Ministry of Defense. Grant sponsor: NIH
support for the CESG Structural Genomics Consortium.
*Correspondence to: Michal Linial, Department of Biological Chem-
istry, Institute of Life Sciences, Hebrew University, Jerusalem 91904,
Israel. E-mail: michall@cc.huji.ac.il
The Supplementary Materials referred to in this article can be found
at: http://www.protonet.cs.huji.ac.il/homologues/
Received 23 February 2004; Accepted 13 May 2004
Published online6 August
(www.interscience.wiley.com). DOI: 10.1002/prot.20235
2004 inWileyInterScience
PROTEINS: Structure, Function, and Bioinformatics 57:531–538 (2004)
© 2004 WILEY-LISS, INC.

Page 2

all homologies when fixing the rate of false positives to
1/1000.8In the case that the sequence similarity is below
the level of secure inference (often referred to as the
“twilight zone”), finding the protein family borders and the
evolutionary remote homologues of a given protein is
desirable. One should bear in mind that naive sequence-
based search methods that are based on similarity scoring
are often inappropriate for multidomain proteins and for
proteins that are a result of convergent evolution.
The inherent limitation of detecting remote homologues
was estimated based on a structural benchmark.9Several
methodologies for navigating in the protein space for a
better detection of remote homologues were proposed.
These include sequence hopping,10transitivity of homol-
ogy,11intermediate sequences,1tuned iterative–profile,12
profile–profile,13jumping alignments,14string-based ker-
nels,15and some hybrid methods.16
Herein we suggest that a construction of the protein
sequence space into a connected graph that captures the
trace of evolutionary relatedness is useful in finding
remote homologues. This method achieves satisfactory
results that are comparable with results that rely on
additional structural and functional knowledge. We sug-
gest that the scaffold of the protein space as presented by
ProtoNet17can be used to infer structural and functional
information for remote homologues with high reliability.
The success in inferring structural and functional related-
ness for selected families by the ProtoNet navigation
method is discussed.
METHODOLOGY
Methods for Detecting Remote Homologues
A comparative study that tested the success in detecting
remote homologues by various published strategies was
published.18The list of protein homologues was manually
selected to represent both structurally and functionally
(i.e., similar enzymatic activities) related families of re-
motehomology.Theproteinsinthetestsetwerecharacter-
ized by their very low pairwise sequence similarity.
For each homologous family, multiple strategies were
applied and combined with the goal of unifying the pro-
teins that belong to the same evolutionary-related group.
Thestrategiesthatwereusedincludedadvancedsequence-
based search methods, including FASTA walk,19PSI-
BLAST,20HMMer,21MAST,22and PROBE.23Each of the
methods initiated a search from a query protein from the
group of proteins, and the number of proteins from the
group that were identified by the search as remote homo-
logues was noted (Table 1). We used this set as a bench-
mark for testing the ability of the ProtoNet cluster-map of
the protein space to perform such a unification of protein
families. We evaluated and quantified the degree of speci-
ficity and sensitivity in detecting those families.
ProtoNet Graph
We base our study on the database of ProtoNet 2.4
(http://www.protonet.cs.huji.ac.il).17ProtoNet is an auto-
matically generated, hierarchical classification of all pro-
tein sequences from Swissprot (release 40.28 with 114,033
sequences). Each sequence is a vertex in a graph, and each
pair of sequences defines an edge whose weight is the
average pairwise BLAST E-score between those two se-
quences. A pairwise merging process is conducted as part
ofanagglomerativeclustering,wheretheinitialsequences
are used as singleton clusters. The clustering process
follows an order of similarity, where each merge is the
most significant in terms of the lowest average E-score
between two groups of sequences. The main principle
applied in ProtoNet is the extensive use of restricted
transitivity based on hidden intermediate sequences,
coupled with a statistical and biological validation of the
hierarchy of the resulting graph. The resulting structure of
the graph allows one to navigate through it and to evaluate
the capacity of the ProtoNet algorithm to detect remote
homologues (see examples24).
RESULTS AND DISCUSSION
Unifying Protein Families in ProtoNet
ProtoNet is a bottom-up tree construction in which
biologically relevant connections between proteins are
evident.24In the initial steps of the clustering process,
connections between closely related sequences are made
but at more advanced stages of the clustering process,
more questionable connections are included; at the root of
the tree, all proteins are interconnected (except for a few
hundred singletons). The hierarchical structure of the
graph that is built by ProtoNet enables one to follow the
connections between different groups of sequences when
considering the progress of the clustering process. Intu-
itively, one can view it as a reflection of the evolutionary
diverging process. As the clustering process advances,
similar sequences aggregate to form groups that then
unite to form subfamilies. As the clustering algorithm
progresses, the families are enlarged, thus forming super-
families and making remote homologues prevalent. As the
number of pairwise connections becomes sparse during the
progress of the clustering algorithm (and the statistical
TABLEI.AGlobalAnalysisfortheCorrespondenceof
ProtoNetClusterswithStructuraland
FunctionalProperties
Properties
SCOPFamily
All
20–1000
SCOPSuperfamily
All
20–1000
EnzymeEC.4thdigit
All
20–1000
EnzymeEC.3rddigit
All
20–1000
#Best
ClustersCSa
Specificityb
Sensitivityc
1283
968
0.84
0.86
0.96
0.98
0.88
0.88
854
607
0.77
0.78
0.95
0.97
0.79
0.80
2157
1581
0.75
0.76
0.88
0.88
0.84
0.85
206
150
0.50
0.50
0.87
0.89
0.55
0.55
aCorrespondence score.
bSpecificity ? TP/(TP ? FP).
cSensitivity ? TP/(TP ? FN). See text for definitions.
532
O. SHACHAR AND M. LINIAL

Page 3

significance of the pairwise connections are extremely low;
i.e., BLAST E-score as low as 100), the representation of
ProtoNet resembles a graph in which the edges connect
groups of sequences (clusters) rather than individual
sequences.
The properties of the graph in view of its biological
content were evaluated by comparing each cluster to other
classification methods such as InterPro,25SCOP,26EN-
ZYMES,27and other external annotation sources. To
evaluate the quality of ProtoNet clusters for structural
and functional properties, we scored all best clusters. We
define such a score as a correspondence score (CS). The CS
for a certain cluster and a given keyword (i.e., Enolase
N-terminal domain-like family, in SCOP) measures the
correlation between the cluster and that keyword, using
the intersect–union ratio.
Correspondence Score (cluster C for keyword K) ? ?c?k?/
?c?k? ? TP/(TP ? FP ? FN), where c is the set of annotated
proteinsinclusterC,kisthesetofproteinsannotatedwith
K, TP, FP, and FN standing for true positives, false
positives, and false negatives, respectively. TP ? the
number of proteins in cluster C that have keyword annota-
tion K; FP ? the number of annotated proteins in cluster C
that do not have keyword annotation K; FN ? the number
of proteins not in cluster C that have keyword annotation
K. The cluster receiving the maximal CS for keyword K is
considered the cluster that best represents K within the
ProtoNet tree, and is called the best cluster. The score for a
given cluster on keyword K ranges from 0 (no correspon-
dence) to 1 (the cluster contains exactly all of the proteins
with keyword K; maximal correspondence to the keyword).
For annotation keywords related to structural proper-
ties, we used all keywords based on SCOP (fold, superfam-
ily, family, and domain levels), and for a functional view,
we used the ENZYME (4 levels of EC hierarchy) and GO
(in 3 categories—molecular function, cellular process, and
cellular localization) annotations. A complete list of all
annotations with the associated ProtoNet best clusters is
available (www.protonet.cs.huji.ac.il/best_cluster/).
The results from such analysis confirm that clusters
that perfectly match functional and structural annotations
as defined by the above-listed sources are abundant within
the ProtoNet graph. Representative results are shown in
Table I. The results for the SCOP family level (contains
1283 keywords) is CS ? 0.84 (with a higher score for
proteins of a single domain; see Discussion section). By
limiting the analysis for cluster size ranges from 20 to
1000 proteins, the CS is somewhat higher (0.86). The
average CS for best clusters for the E.C. ENZYME fourth
level in hierarchy (contains 2157 keywords) is quite high,
but drops significantly for the third digit of the EC
classification.
Results from the best cluster view point to a relatively
narrow interval along the clustering process, where the
connection between protein families starts to dominate.
Thus, two protein families that share the same cluster in
the cluster-map described at that interval may be viewed
as remote homologues. This last principle guided us when
we tested for the ability of ProtoNet system to suggest
remote homologues.
To compare the ability of ProtoNet to detect remote
homologues to that of other available search methods, we
choose to test a manually selected set of remote homo-
logues as presented in a comparative survey.18In Table II,
examples of proteins that share structural and functional
relatedness are listed. We duplicated the comparative
analysis to include ProtoNet method as an additional,
testable methodology in view of the previously described 5
state-of-the-art methods for remote homologues detection.
The remote family representatives are marked by the
characteristics of the relevant protein functional family
(i.e., Winged helix-turn-helix (HTH) DNA-binding do-
mains; Table II). We began each search with the query
protein listed in Table II and “climbed” the cluster hierar-
chy until reaching the largest cluster in which the target
proteins that are included in the defined functional family
are encountered. The process is aborted, and no success is
recorded when the cluster that includes the query and the
target proteins annotate less than 85% of the proteins in
the cluster (denoted as TP—true positives; unannotated
proteins are not considered for such procedure). The
number of protein families and representatives that were
united under that search was reported. In the case that the
selected largest cluster is already contaminated by unre-
lated identifiers or keywords, we marked it (by minus sign;
Table II) to indicate a failure in the specificity (purity) of
the cluster.
Evaluating the Success in Detecting Remote
Homologues
Inspection of the results from Table II indicates that our
simple procedure using the scaffold of ProtoNet clusters
performs as well as the best method out of the 5 that were
tested for almost every query. Note that the benchmark
was completely independent of our work and was selected
to evaluate the performance of the available 5 independent
methods for detecting remote homologues. The different
methods are based on exhaustive dynamic programming
(FASTA), iterative profile search (PSI-BLAST, PROBE),
and traditional profile-based searches (GRIBSKOV, HM-
Mer). For details on the listed methods, see ref. 18).
A strong characteristic of our method is its symmetry in
the results between query and target protein. This is
simply due to the symmetry of the search up a cluster
tree—the cluster that unites a query A with a target B is
the same cluster that unites B with A when B is the query
and A is the target. This symmetry property implies
consistency of the search by the ProtoNet method. It can be
appreciated from Table II in that the other methods do not
perform symmetrically.
Othermaincharacteristicsofourmethodarethesimplic-
ity and speed of the search, which requires only basic
ProtoNet navigation skills of the user, and its robustness,
as the cluster-map platform already contains all sequences
from Swissprot (ver. 40.28). Every sequence can be re-
ferred to as a query or target sequence for such a search.
REMOTE HOMOLOGUES IN PROTEIN FAMILIES
533

Page 4

TABLEII.ComparingtheSuccessinDetectingRemoteHomologuesbyAlternativeMethods
Functionalfamily
WingedHTHDNA-binding
domains(1opc,1aoy,1dprB,
1hst,1ecl,1bgw,1smtA,1lea,
1fokA)
Querysequence
1.ompr_ecoli(1opc)
A
2
B
1
C
1
D
1
E
2
PN
2?
Cluster
(no.of
proteins)
226420(1088)
Remarks
LowTP,Detected:
1b9w
2.argr_ecoli
3.1lea
1
/
1
/
1
/
1
/
1
/
1
3226170(348)
TP?1forSF
Detected:1smtA,
1dprB
Nucleicacid–bindingproteinswith
OB-fold
1.S1domains,translation
initiationfactorIF1(1ah9)
2.Coldshockprotein(1mjc)
(1asyA,1lylA,1cuk,3ullA,
1cmkA,1pfsA)
1.1ah9111111
2.cspa_ecoli111101
3.1asy
/////2210265(143)Detected:1lyl
1.Pertusistoxin
2.Aerolysin(domain)
3.LINKdomains
4.CTLdomain
1.tox3_borpe
2.lems_human
11
1
0
0
1
1
0
1
1
1 Ex
1.Chromodomain
2.DNA-bindingproteins(1sap)
mod3_human
1sap
1
1
1
1
1
1
1
1
1
1
1
1
1.Miteallergen
2.Immunoglobulin-like
def2_derfa111111
Urease**ure1_kleae 4[?6]7 1013 10 225659(253)
TP??86%
1.PAPSreductase
2.ATPsulphurylase
3.N-typeATPPpases(1nsyA,
1gpmA)
1.mt16_yeast132233?
226597(841)LowTPforallthree;
highTPfor1and
2incluster
212193
2.nade_bacsu13311
3.cysd_ecoli33333
1.HIT(4rhn)
2.Ap4Ahydrolase
3.GaIT(1hxpA)
1.fhit_human 2[?1]33312221320(48) CouldnotfindAp4A
proteinsthrough
search.HighTP
2.gal7_ecoli11311
1.CovalentNAD-bindingoxidases
(1ahu)
2.UDP-N-
acetylenolpyruvoylglucosamine
(2mbr)
1.1ahu122012225895(79)HighTP
2.murb_ecoli1[?1]1120
CoenzymeAtransferase
1.Glutaconate
2.Succinyl,acetate,3-oxoadipate,
butyrate,etc.
1.X81440
2.atoa_ecoli
2
2
2
2
2
1
2
1
2
1
/ 225850(51)X81440notpresent
inSwissprot.2nd
queryunifiedthis
family
2(*)
1.Dioxygenase(1han)1.bphc_burce122122 222839(70)
TP?1forSF
Includingalso
1bylfromthisSF
2.Glyoxylase(1froA) 2.igul_human12221
Functional families that unify two or more sequence families were combined, as in Holm.18The success in detection is marked by the number of
successes in view of the predetermined representatives for remote homology. Alternative methods for remote homologue detection were applied
over single query proteins, and an enumeration of the sequence families that were recognized by each of the methods was recorded (a result of 1
means the query is recovered; zero indicates failure to identify it). A search in ProtoNet tree was performed in the same way by climbing the
cluster hierarchy, starting from a query protein and determining the cluster in which the protein identifier of the target proteins reached ? 85%
TP. The methods applied are abbreviated as follows: A. FASTA walk; B. PSI-BLAST; C. PROBE; D. GRIBSKOV; and E. HMMer (according to
Holm18); PN, ProtoNet. The mark (?) indicates contamination by unrelated structural families. The mark (/) indicates a query examined solely by
ProtoNet. Ex, the number of protein exploded. (*), the entry X81440 could not be matched in ProtoNet s proteins, so we indicated the smallest
cluster that unites both families of that query. (**), discussed in detail in the text.
534
O. SHACHAR AND M. LINIAL

Page 5

While performing our searches, we were able to find in
some functional families more examples of remote homo-
logues that were not queried in the original benchmark
(marked by underline, Table II). These examples were
found as a by-product of the search in ProtoNet, and were
added as additional representatives to enrich the list in
Table II.
The protein family of the urease-related hydrolases
serves to illustrate the traces of diversifying proteins
throughout evolution. Several methods were applied to
unify the members of the urease family.28All methods
usedstructuralalignment,sequencealignment,andprofile-
based search.10The entire group, as defined in the original
study,10contains 12 protein families that share structur-
ally similar active-site architecture. Of those proteins, 9
are members of the metallo-dependent hydrolases SCOP
superfamily, 2 are members of a related superfamily
named composite domain of metallo-dependent hydrolases,
and an additional one belongs to a different superfamily
(arylphosphatase). Of the hydrolases that are in the data-
base, ProtoNet defined 10 out of 12 in a relatively pure
cluster (A225659, 253 proteins). This cluster best repre-
sents the metallo-dependent hydrolases superfamily hier-
archy.
Inspecting this cluster obtained in ProtoNet shows that
it includes all previously described urease-related super-
family proteins, as well as additional yet undefined related
members, as shown in Table III. The different representa-
tives in this large family all share an enzymatic activity
marked by EC 3.5.-.-, with the exception of proteins having
a different catalytic activity (EC 3.8.-.-.). Interestingly, the
evolutionary and functional relatedness between those
two apparently different activities was confirmed.29Addi-
tionally, not only is the high coverage of the urease-related
superfamily (in cluster A225659) validated, but monitor-
ing the descendant clusters in the cluster-map provides a
good description of the partition of this functional family
into more refined subfamilies. Relationships between the
subfamilies can be proposed, and the divergence process of
the family is traceable.
Figure1showsthepairwisesimilaritywithinproteinsof
the cluster A225659 (253 proteins) as a matrix reflecting
all-against-all pairwise similarity BLAST E-scores. Dark-
est color indicates a maximal BLAST E-score (E-score ?
0). The gradient in color intensity indicates a decrease in
the similarity score. White indicates an E-score that is
worse than 100. The domination of the white color reflects
the very low pairwise similarity among proteins in the
cluster. Among all possible pairwise connections, ?70%
are worse than E-score 100, indicating the sequence
remoteness resulting from the BLAST search. The order of
the proteins in the matrix reflects the construction of the
ProtoNet tree for this subgraph (starting at the bottom
right toward the top left). An interactive illustration
(color-coded) of the weak connections between all 253
proteins in this cluster is provided at www.protonet.cs.hu-
ji.ac.il/homologues/.
In addition to the examples described in Table III and in
Figure 1, additional remote homologues are easily de-
tected. For example, the proteins that share the histone
core proteins H2A, H2B, H3, and H4 are linked together in
a subtree of ProtoNet with a variety of transcription
factors, such as the transcriptional initiator TFIIB, TFIID,
TATA box, CAAC box binding protein, and more. Despite
extremely low sequence similarity (cluster A224843, 284
proteins), all these proteins share an identical structural
fold.30
In another very interesting case, similarity between
alcohol dehydrogenases and crystalline of the eye and
some proteins in synaptic vesicles in the nerve system has
been proposed.31Such homology is apparent at the se-
quence similarity level only through some weak connect-
ing intermediate sequences. Proteins of the alcohol-
dehydrogenases(alsopolyol-,threonine-,archaeonglucose-
dehydrogenases), eye lens zeta-crystallins, Escherichia
coli quinone oxidoreductase, synaptic vesicle VAT-1 pro-
teins, enoyl reductases of mammalian fatty acid, and yeast
erythronolide synthases share the same structural fold,
with only very few conserved amino acids throughout this
diverged superfamily.32This superfamily was termed me-
dium-chain dehydrogenases/reductases (MDRs). A cluster
in ProtoNet with 226 proteins (cluster A220748) combines
all the MDR proteins that carry very different functional
roles but nevertheless share a high degree of structural
andbiophysicalproperties.Byusingtheinteractivevisual-
ization tools within ProtoNet (http://www.protonet.cs.
huji.ac.il/), the connectivity among the proteins of the
MDR superfamily is shown. Note that the intermediate
sequences that connect specific families are traceable by
inspecting the pairwise matrix (as in Fig. 1, and see
www.protonet.cs.huji.ac.il/homologues/).
The ProtoNet tree allows for the testing of the connectiv-
ity among proteins, thus detecting remote homologues.
However, our examples are based on analyzing proteins
from the Swissprot database (?114,000 proteins). As the
number of protein sequences currently available is at least
10 times higher, the relevance and the robustness of our
method toward the larger database should be re-evalu-
ated. We developed an extended version of ProtoNet that
includes over 1 million proteins (ProtoNet 3.0), based on
Swissprot 41.12 combined with TrEMBL 24.8. Many of the
proteins in this combined data set are still poorly anno-
tated. In ProtoNet 3.0, nodes in the tree that are candi-
dates for representing clusters of remote homologues are
eventually much larger. We tested the robustness of the
results in the extended version of ProtoNet. In general,
despite a ?9-fold increase in the database size and a
reduced quantity and quality of the associated annota-
tions, the performance of the system is only slightly
reduced. We illustrate it for the urease superfamily that
was discussed above. A cluster representing the urease-
related superfamily (A293766, ProtoNet 3.0) is composed
of 1508 proteins that include 16 out of the 17 groups (as in
Table III) with a very high sensitivity (i.e., collecting all
proteins from the database that belong to this enzymatic
group). One additional urease-related enzymatic group
(N-acyl-D-amino acid deacylase) that was missed from the
Swissprot based cluster (Table II) is detected in cluster
REMOTE HOMOLOGUES IN PROTEIN FAMILIES
535

Page 6

A293766. This example illustrates the potential of the
ProtoNet tree to detect remote homologues even in a
relatively poorly annotated database. It is worth noticing
that among the 1508 proteins, most of them (58%) have no
Prosite annotation, and 21.1% are marked as hypothetical
proteins. Thus, the inference of functional groups is a
direct byproduct of navigating the ProtoNet tree.
A careful inspection of the families with remote homo-
logues showed that the performance of ProtoNet is very
satisfactory for single-domain proteins (as in the case of
histones, urease, and MDR proteins); a much lower suc-
cess rate can be obtained for proteins that are multido-
main. The relatively low scores in the best cluster analysis
(www.protonet.cs.huji.ac.il/best_cluster/) best reflect such
instances. While ProtoNet is based on a whole protein
clustering, and remote homology (especially from a struc-
tural perspective) is often associates with a domain level,
it is appropriate to use this methodology with a critical
objective.
In summary, we offer the use of ProtoNet’s cluster-map
as a platform for searching for remote homologues and for
identificationofweakconnectionsbetweenfunctionalfami-
lies. We propose a simple search procedure for finding
remote homologues for a query sequence and compare it to
Fig. 1.
is shown in the callouts. The matrix shows all 253 proteins in cluster A225659. Black indicates maximal BLAST E-score (0). Dark gray indicates very
significant pairwide similarity. A gradient of black to white parallels a decrease in the similarity score.
All-against-all BLAST E-score in a subtree of ProtoNet for urease-related proteins. Representative enzymatic activity of some of the proteins
536
O. SHACHAR AND M. LINIAL

Page 7

a benchmark of performance for different methods that
rely on information beyond the elementary pairwise simi-
larity. We show that in almost all cases, our search method
performs as well as the best method. Our method is very
fast and consistent due to the fact that it relies on a
predetermined scaffold of the protein space that is in
accord with the process of evolutionary diversification.
Partitioning the protein space into homologous families
ismotivatedbyStructuralandFunctionalGenomicsinitia-
tives. As seen from the benchmark and the additional
examples, the identification of remote homologues for a
given sequence extends our ability in predicting structural
relatedness and in many instances also the functional
characteristics. We propose a robust method for finding
such remote homologues and defining structural and func-
tional superfamilies.
ACKNOWLEDGMENTS
We would like to thank the ProtoNet team, and espe-
cially Uri Inbar and Hillel Fleischer, for excellent database
management. We thank Noam Kaplan and Menachem
Fromer for a critical reading and useful suggestions. We
thank SCCB, the Sudarsky Center for Computational
Biology for a fellowship support (O.S.).
REFERENCES
1. Sternberg MJ, Park J, Karplus K, Barrett C, Hughey R, Haussler
D, Hubbard T, Chothia C. Sequence comparisons using multiple
sequences detect three times as many remote homologues as
pairwise methods. J Mol Biol 1998;284:1201–1210.
2. Russell RB, Saqi MA, Sayle RA, Bates PA, Sternberg MJ. Recogni-
tion of analogous and homologous protein folds: analysis of
sequence and structure conservation. J Mol Biol 1997;269:423–
439.
3. Stark A, Sunyaev S, Russell RB. A model for statistical signifi-
cance of local similarities in structure. J Mol Biol 2003;326:1307–
1316.
4. Saier MH Jr. Phylogenetic approaches to the identification and
characterization of protein families and superfamilies. Microb
Comp Genomics 1996;1:129–150.
5. Rost B. Enzyme function less conserved than anticipated. J Mol
Biol 2002;318:595–608.
6. Rost B. Twilight zone of protein sequence alignments. Protein Eng
1999;12:85–94.
7. Jones DT, Swindells MB. Getting the most from PSI-BLAST.
Trends Biochem Sci 2002;27:161–164.
8. Karwath A, King RD. Homology induction: the use of machine
learning to improve sequence similarity searches. BMC Bioinfor-
matics 2002;3:11.
9. Muller A, MacCallum RM, Sternberg MJ. Benchmarking PSI-
BLAST in genome annotation. J Mol Biol 1999;293:1257–1271.
10. Holm L, Sander C. An evolutionary treasure: unification of a broad
set of amidohydrolases related to urease. Proteins 1997;28:72–82.
11. BrennerSE,ChothiaC,HubbardTJ.Assessingsequencecompari-
sonmethodswithreliablestructurallyidentifieddistantevolution-
ary relationships. Proc Natl Acad Sci USA 1998;95:6073–6078.
12. Schaffer AA, Aravind L, Madden TL, Shavirin S, Spouge JL, Wolf
YI, Koonin EV, Altschul SF. Improving the accuracy of PSI-
BLAST protein database searches with composition-based statis-
tics and other refinements. Nucleic Acids Res 2001;29:2994–3005.
13. Yona G, Levitt M. (2002). Within the twilight zone: a sensitive
profile–profile comparison tool based on information theory. J Mol
Biol 2002;315:1257–1275.
14. Spang R, Rehmsmeier M, Stoye J. A novel approach to remote
homology detection: jumping alignments. J Comput Biol 2002;9:
747–760.
15. Leslie CS, Eskin E, Cohen A, Weston J, Noble WS. Mismatch
string kernels for discriminative protein classification. Bioinfor-
matics 2004;20:467–476.
16. Jaroszewski L, Rychlewski L, Godzik A. Improving the quality of
twilight-zone alignments. Protein Sci 2000;9:1487–1496.
17. Sasson O, Vaaknin A, Fleischer H, Portugaly E, Bilu Y, Linial N,
Linial M. ProtoNet: hierarchical classification of the protein space.
Nucleic Acids Res 2003;31:348–352.
18. Holm L. Unification of protein families. Curr Opin Struct Biol
1998;8:372–379.
19. Pearson WR. Using the FASTA program to search protein and
DNA sequence databases. Methods Mol Biol 1994;24:307–331.
20. Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller
W, Lipman DJ. Gapped BLAST and PSI-BLAST: a new generation
of protein database search programs. Nucleic Acids Res 1997;25:
3389–3402.
TABLEIII.TheEnzymaticGroupsAssociatedwithProtoNetClusterA225659
EnzymeName
(1)Adenosinedeaminase
(2)Adeninedeaminase
(3)Allantoinase
(4)AMPdeaminase
(5)N-isopropylammelideisopropylamidohydrolase
(6)Atrazinechlorohydrolase
(7)Hydroxydechloroatrazineethylamonohydrolase
(8)Cytosinedeaminase
(9)Dihydropyrimidinase2
(10)Guaninedeaminase
(11)Imidazolonepropionase
(12)N-acetylglucosamine-6-phosphatedeacetylase
(13)N-acyl-D-aspartatedeacylase
(14) D-aminoacylase
(15)N-acyl-D-glutamatedeacylase
(16)Dihydroorotase(DHOase)
(17)Urease(Ureaamidohydrolase)
no.ofProteinsinthe
Cluster/no.inDatabase
28/28
6/6
6/6
11/11
1/1
1/1
1/1
1/3
10/10
7/8
23/23
6/6
1/1
1/1
1/1
49/49
78/78
E.C.
3.5.4.4
3.5.4.2
3.5.2.5
3.5.4.6
3.5.99.4
3.8.1.8
3.5.99.3
3.5.4.1
3.5.2.2
3.5.4.3
3.5.2.7
3.5.1.25
3.5.1.83
3.5.1.81
3.5.1.82
3.5.2.3
3.5.1.5
The annotations for 17 different groups are extracted from the ENZYME database. Note that for
almost all groups, the coverage is complete, with no additional proteins carrying the same
enzymatic activity reported in the database used.
REMOTE HOMOLOGUES IN PROTEIN FAMILIES
537

Page 8

21. Karplus K, Barrett C, Hughey R. Hidden Markov models for detect-
ingremoteproteinhomologies.Bioinformatics1998;14:846–856.
22. Bailey TL, Gribskov M. Methods and statistics for combining
motif match scores. J Comput Biol 1998;5:211–221.
23. Neuwald AF, Liu JS, Lipman DJ, Lawrence CE. Extracting
protein alignment models from the sequence database. Nucleic
Acids Res 1997;25:1665–1677.
24. Sasson O, Linial N, Linial M. The metric space of proteins—
comparative study of clustering algorithms. Bioinformatics 2002;
18(Suppl 1):S14–S21.
25. Mulder NJ, Apweiler R, Attwood TK, Bairoch A, Bateman A,
Binns D, Biswas M, Bradley P, Bork P, Bucher P, Copley R,
Courcelle E, Durbin R, Falquet L, Fleischmann W, Gouzy J,
Griffith-Jones S, Haft D, Hermjakob H, Hulo N, Kahn D, Kanapin
A, Krestyaninova M, Lopez R, Letunic I, Orchard S, Pagni M,
Peyruc D, Ponting CP, Servant F, Sigrist CJ. InterPro: an
integrated documentation resource for protein families, domains
and functional sites. Brief Bioinform 2002;3:225–235.
26. Lo Conte L, Ailey B, Hubbard TJ, Brenner SE, Murzin AG,
Chothia C. SCOP: a structural classification of proteins database.
Nucleic Acids Res 2000;28:257–259.
27. Bairoch A. The ENZYME database in 2000. Nucleic Acids Res
2000;28:304–305.
28. Heger A, Holm L. Sensitive pattern discovery with “fuzzy” align-
ments of distantly related proteins. Bioinformatics 2003;19(Suppl
1):I130–I137.
29. Sadowsky MJ, Tong Z, de Souza M, Wackett LP. AtzC is a new
member of the amidohydrolase protein superfamily and is homolo-
goustootheratrazine-metabolizingenzymes.JBacteriol1998;180:
152–158.
30. Baxevanis AD, Landsman D. Histone Sequence Database: new
histone fold family members. Nucleic Acids Res 1998;26:372–375.
31. Linial M, Levius O. VAT-1 from Torpedo is a membranous
homologue of zeta crystallin. FEBS Lett 1993;315:91–94.
32. Persson B, Zigler JS Jr, Jornvall H. A super-family of medium-
chaindehydrogenases/reductases(MDR):sub-linesincludingzeta-
crystallin,alcoholandpolyoldehydrogenases,quinoneoxidoreduc-
tase enoyl reductases, VAT-1 and other proteins. Eur J Biochem
1994;226:15–22.
538
O. SHACHAR AND M. LINIAL