Full text of DOI:10.1093/bioinformatics/17.10.920

Vol. 17 no. 10 2001
Pages 920–926
BIOINFORMATICS
Automatic rule generation for protein annotation
with the C4.5 data mining algorithm applied on
SWISS-PROT
Ernst Kretschmann, Wolfgang Fleischmann and Rolf Apweiler
The EMBL Outstation, The European Bioinformatics Institute, Wellcome Trust
Genome Campus, Hinxton, Cambridge CB10 1SD, UK
Received on April 20, 2001; revised and accepted on July 8, 2001
ABSTRACT
found), the literature (in which it was mentioned), etc.
Motivation: The gap between the amount of newly
submitted protein data and reliable functional annotation in
public databases is growing. Traditional manual annotation
by literature curation and sequence analysis tools without
the use of automated annotation systems is not able
to keep up with the ever increasing quantity of data
that is submitted. Automated supplements to manually
curated databases such as TrEMBL or GenPept cover raw
data but provide only limited annotation. To improve this
situation automatic tools are needed that support manual
annotation, automatically increase the amount of reliable
information and help to detect inconsistencies in manually
generated annotations.
• Information or annotation: the statement of an aspect
that is relevant or important to describe the protein
as a whole or parts of it, e.g. ‘It is expressed in the
mitochondrion’, ‘Amino acids 1–26 encode a Signal’,
etc.
• Knowledge: the process that draws conclusions about
an unknown protein using gathered information,
e.g. ‘The protein sequence contains pattern x. Since
all known sequences having this pattern belong
to transmembrane proteins, this should also be a
transmembrane protein.’
• Data mining: any technique that uses information to
gain knowledge on data.
Results: A standard data mining algorithm was suc-
cessfully applied to gain knowledge about the Keyword
annotation in SWISS-PROT. 11 306 rules were generated,
which are provided in a database and can be applied to
yet unannotated protein sequences and viewed using a
web browser. They rely on the taxonomy of the organism,
in which the protein was found and on signature matches
of its sequence. The statistical evaluation of the generated
rules by cross-validation suggests that by applying them
on arbitrary proteins 33% of their keyword annotation can
be generated with an error rate of 1.5%. The coverage
rate of the keyword annotation can be increased to 60%
by tolerating a higher error rate of 5%.
INTRODUCTION
How to obtain information about a protein? If the protein
was biochemically characterized before and this informa-
tion was entered into a database like SWISS-PROT, which
is a completely human expert controlled and maintained
database (Bairoch and Apweiler, 2000), one can simply
make use of the provided information, i.e. a human be-
ing has used his knowledge to compose annotations on
this very protein data and established a one to one rela-
tionship between this data and its annotation which can be
used by others. However, often the information is incom-
plete, which is a fact for the majority of the known pro-
teins. Many of those poorly annotated proteins are stored
in databases like TrEMBL (Bairoch and Apweiler, 2000),
which is only partly annotated by human experts but also
by automated annotation systems like EDIT to TrEMBL
(Mo¨ller et al., 1999) and RuleBase (Fleischmann et al.,
1999; Apweiler, 2001). The protein can even be hypothet-
ical, so there is no information available at all.
Availability: The results of the automatic data mining
process can be browsed on http://golgi.ebi.ac.uk:8080/
Spearmint/ Source code is available upon request.
Contact: kretsch@ebi.ac.uk
TERMINOLOGY
This paper is about data, information, and knowledge on
protein sequences. As far as we know there is no standard
definition to distinguish between these concepts. In the
following, we are going to use the definitions given below:
In these cases one mostly resorts to sequence similarity
or signature searches, hoping to find well annotated
protein features sharing some similarity with the protein
in question. Apart from similarity searches against com-
prehensive, non-redundant protein sequence databases
• Data: the measurable or observable facts, e.g. the
sequence, the organism (in which the protein was
c
920
ꢀ Oxford University Press 2001

Automatic rule generation for protein annotation
like SWISS-PROT and TrEMBL (Apweiler, 2000), the
use of protein sequence signature databases such as
Prosite (Hofmann et al., 1999), PRINTS (Attwood et
al., 2000) or Pfam (Bateman et al., 2000) can be helpful
as are protein cluster databases like SYSTERS (Krause
et al., 1999) and CluSTr (Kriventseva et al., 2001). In
those cases, there is a many-to-many relationship between
annotation and data, i.e. one annotation is stored for many
proteins and one protein sequence might match various
signatures and their annotation. Obviously, the process of
gathering, analyzing, evaluating, and deriving information
is time-consuming and cumbersome. It can be regarded as
manual data mining across various databases.
We have developed a method to automate this process
for a subset of the information available in SWISS-PROT,
the Keyword Line. Keywords are particularly useful for
analysis because they are controlled, limited in number
(at the time of this writing there were 850 different
Keywords allowed), they show little inherent structure or
dependencies and are either annotated or not. These facts
make automated knowledge acquisition much easier as for
comment lines and description lines, which often are in
unstructured free text.
Fig. 1. Example of data distribution in InterPro IPR003009 (only
part of data is shown). The first column contains the SWISS-PROT
accession numbers of some proteins in this entry.
Mammal?
yes
no
do nothing
Prosite
(5 instances)
PS00487?
yes
no
The implementation uses the C4.5 data mining algo-
rithm to detect decision trees which are an equivalent no-
tation to rules. C4.5 shows particularly good results for
non noisy data, which is the case for SWISS-PROT. The
derived rules are not only fitting the training set, but are
also human readable and kept short. This is obtained by
an elaborated heuristic approach inherent to the standard
algorithm. Also statistical evidence is given for every rule,
which can be used to order rules in terms of confidence.
This property can be used to select subsets of rules for dif-
ferent applications, i.e. only the highly confident ones for
error critical purposes where coverage is less important
and all of the generated rules where coverage is the main
concern.
do nothing
(1 Instance)
annotate ’FAD’
(3 instances)
Fig. 2. Decision tree describing the data in Figure 1.
Prosite pattern PS00487, three to Pfam pattern PF01493,
five belong to mammalia and three have the Keyword
‘FAD’. The distribution is as follows.
A decision tree is generated that has a preferably small
number of leaves to make rules better readable and at the
same time more reliable. Less leaves mean that on the
average there are more examples per leaf that give the rule
better statistical confirmation. In general, there are several
possible equivalent decision trees. The example decision
tree in Figure 2 covers all the instances in the training set
in Figure 1.
SYSTEM AND METHODS
Algorithm
But the decision tree in Figure 3 classifies the data
more compactly. The problem of finding the optimal
decision tree is known to be NP-complete (Hyafil and
Rivest, 1976). C4.5 uses the gain ratio criterion, which
is based on information theory and produces suboptimal
trees heuristically (Quinlan, 1993). Note that if there are
two instances having the same core data but a different
annotation, there is no tree that classifies all examples of a
training set correctly.
The precision of a tree can be checked by analyzing
the number of correct and incorrect classifications it
produces when applied on the training set. This analysis
gives the number of True Positives (TPs) (annotation
One of the basic ideas of artificial intelligence algorithms
is to derive knowledge from training sets and apply it on
yet unknown data. The C4.5 algorithm expects input in a
tabular format where the last column contains the target,
in this particular case the information if a given keyword
is present or not. The previous columns store core data
about the proteins like taxonomy details or the presence
of sequence patterns. The algorithm tries to derive the
contents of the last column by using the information in
the other columns.
To illustrate the procedure, a simple example for the
SWISS-PROT proteins in InterPro (Apweiler et al., 2000)
IPR003009 is given. One of those proteins matches to
921

E.Kretschmann et al.
estimation has to be performed of how well they will
perform on unknown data in terms of coverage and error
rate. This aspect was tested by a tenfold cross-validation
(see Results).
Pfam
PF01493
yes
no
The standard algorithm was designed to classify in-
stances into groups where there is an interest in all classes.
Yet in this particular application there is no need for rules
suggesting the non-annotation of certain Keywords. The
standard statistical evaluation implemented in C4.5 was
tried as a method to order rules in terms of quality.
Parameters to trigger the calculation were adapted to the
particular problem but the results were unsatisfactory.
Therefore a procedure was chosen that derives confidence
by exclusively using the number of TP and FP examples.
The formula calculates the following value of likelihood:
given the number of TP and FP examples it calculates,
which rules lie above a given threshold in 95% of all
cases. To illustrate the idea: suppose drawing from an
urn containing an infinite number of balls. Drawing ten
black balls and one white ball, over which value does the
true ratio black to white balls in the urn lie in 95% of
the cases? (TP = True Positives, FP = False Positives,
c = confidence)
annotate ’FAD’
(3 instances)
do nothing
(6 instances)
Fig. 3. Compact decision tree describing the data in Figure 1.
exists in instance and is predicted), True Negatives (TNs)
(annotation does not exist in instance and is not predicted),
False Positives (FPs) (annotation does not exist in instance
but is predicted) and False Negatives (FNs) (annotation
exists in instance but is not predicted).
A brute force implementation of the C4.5 algorithm
could successively produce a decision tree for every
allowed Keyword using all the protein core data available
in SWISS-PROT. This procedure would produce huge
data tables which can not be analyzed efficiently (every
table would consist of more than 90 000 rows, one for
each protein in SWISS-PROT). To produce decision trees
with a satisfactory confidence fast enough, the number
of instances for this application should be between 100
and 1000. Hence, a subdivision of SWISS-PROT into
protein groups, which ideally contain similar proteins
has to be performed. Thus, the grouping into proteins
common to InterPro (Apweiler et al., 2000) entries will be
analyzed, since those entries usually contain a convenient
number of similar proteins. Other groupings like using
proteins common to CluSTr (Kriventseva et al., 2001)
entries were performed but not analyzed in detail. Brief
investigations of some decision trees starting from CluSTr
entries showed similar results to that starting from InterPro
entries.
z = 1.96
(constant for 95%)
n = TP + FP
TP
p = precision =
TP + FP
ꢀ
p2
n
z2
z2
p
n
z2
n
p + − z ∗
−
+
4n2
2n
c = confidence =
.
1 +
Formula 1. Ordering rules in terms of confidence. The
formula depends on TP and FP examples exclusively.
Confidence gives the value above which all experiments
would lie when an urn experiment was perfomed with the
same distribution of correct and false outcomes.
Figure 4 gives a short overlook, for which confidence
would be calculated for given numbers of TP and FP
examples. To be introduced in the database, a rule had to
have a confidence of 50% or more.
Extensions
The algorithm produces a large amount of rules with
varying qualities. In many cases the annotation is not a
result of sequence signature or taxonomy and therefore a
decision tree trying to classify instances on this basis will
produce annotation at random. The application of those
trees on unknown data would lead to a massive error rate.
Therefore, a selection of the more trustworthy rules has to
be made and evaluated.
There are two steps of statistical evaluation of the
results: firstly, not every generated rule is suitable to be
applied, since many proved to have either a too high
ratio of FPs to TPs or simply too few sample cases to
derive a good statistical confirmation. Therefore, a smart
criterion had to be used to select only the best rules
with a reasonably high confidence. Secondly, once rules
with a reliability over a given threshold are selected, an
IMPLEMENTATION
The core application is Java based and uses the Weka
Machine Learning Software package which is open
source software and issued under the GNU General Public
License (download at http://www.cs.waikato.ac.nz/∼ml/
weka/). The system is divided into a loader module that
translates the core information stored in various databases
into the tabular input format of the algorithm and an
analyzer module that derives rules and stores the result
into a newly created database to allow quick and easy ac-
cess. Thus, the classical pipeline input–processing–output
922

Automatic rule generation for protein annotation
1.0
0.8
0.6
0.4
0.2
0.0
IPRxxx
IPRyyy
SWISS-PROT
IPRzzz
0
1
3
5
10
25
50
100
C4.5
Rule
Rule
Rule
KW z
KW a
KW b
0
20
40
60
80
100
True Positives
Confidence > Threshold?
yes no yes no no
Fig. 4. Ratio TP to FP examples and the resulting confidence. The
no
yes
curves are for 0, 1, 3, 5, 10, 25, 50 and 100 FPs from top to bottom.
Spearmint Database
was implemented with the processing and output unit
tied together, an approach that makes extensions and
maintenance easier than that of a single monolithic
application.
Two different modes of operation were implemented:
One produces rules on the basis of all suitable SWISS-
PROT proteins and writes them to a database. The other
performs the cross-validation and evaluates rules without
storing them. The data flow of both applications is shown
in Figures 5 and 6.
A graphical user interface was developed that allows
browsing of the generated information. It has some
functionality implemented that can be valuable for the
work of the professional annotators but also for a broader
range of applications:
IPRzzz
IPRxxx
TrEMBL
IPRyyy
Fig. 5. Dataflow for a hypothetical production run. Rules are
generated from InterPro families in SWISS-PROT, their confidence
is tested against a given threshold (50%) and they are either
discarded or added into the Spearmint database. From there they
can be applied on proteins in TrEMBL. Note that rules starting from
a given InterPro family are applied on the very same family and that
the distribution of the InterPro families is different in SWISS-PROT
and TrEMBL.
• It is possible to input the accession number of a protein
in TrEMBL and get the suggested keyword annotation
together with a confidence for each keyword.
the other. The influence of the bias is difficult to measure
and is further analyzed below.
Within SWISS-PROT there are different levels of data
quality due to a varying degree of experimental verifica-
tion of different characteristics of a protein. Uncertain or
predicted properties are categorized as probable, poten-
tial, putative and hypothetical with decreasing reliability
in that order (Junker et al., 1999; Apweiler, 2001; http:
//ch.expasy.org/cgi-bin/lists?annbioch.txt).
The general characterization status of a protein can
be taken from the Description Line of the entry. The
annotation of hypothetical proteins is usually bare and
incomplete, which makes their usage in training sets
unreasonable, hence they were not used for this purpose
(apart from very few hypothetical proteins which are not
marked as such in the Description Line). Probable and
putative proteins were kept, but will be filtered out in
future versions of the tool, since their annotation has
unknown reliability.
• It is possible to track inconsistencies in SWISS-PROT
by using SWISS-PROT both as a training and a test set.
Sometimes it is not possible to find a rule without FNs
and/or FPs. Those can be examined to validate their
annotation.
• The manual generation of rules in RuleBase can be
supported by proposing rules for a given set of proteins
for manual processing.
RESULTS
Successively applied on the proteins assembled in each
InterPro entry the algorithm generated 11 306 rules whose
reliability was evaluated by a tenfold cross-validation. The
quality of the rules depends on the quality of the data in the
training set on one hand and on the bias between training
data and the data on which rules are going to be applied on
923

E.Kretschmann et al.
Table 1. Fragments included in trainings set, confidence > 90%
SWISS-PROT
IPRxxx
IPRyyy
trailing digit: 0-8
End
No. of
Covered
No. of
errors
% covered
% errors
IPRzzz
digit
keywords
keywords
0
1
2
3
4
5
6
7
8
9
28 225
28 033
27 899
28 040
27 498
28 058
28 247
28 049
27 748
28 129
9 629
9 533
9 579
9 553
9 340
9 609
9 647
9 386
9 380
9 414
214
214
155
172
214
223
171
210
192
206
33.36
33.24
33.78
33.46
33.19
33.45
33.55
32.71
33.11
32.73
2.22
2.24
1.62
1.80
2.29
2.32
1.77
2.24
2.05
2.19
C4.5
Rule
KW a
Rule
KW b
Rule
KW z
Confidence > Threshold?
yes no yes no no
279 926
95 070
1971
33.26
2.07
no
yes
Table 2. Fragments included in trainings set, confidence > 67%
IPRyyy
IPRxxx
IPRzzz
SWISS-PROT
trailing digit: 9
End
digit
No. of
keywords
Covered
keywords
No. of
errors
% covered
% errors
Count inconsistencies
0
1
2
3
4
5
6
7
8
9
28 225
28 033
27 899
28 040
27 498
28 058
28 247
28 049
27 748
28 129
17 456
17 337
17 148
17 348
17 037
17 311
17 457
17 274
16 905
17 279
1 049
966
1 001
907
1 045
1 142
983
1 087
998
58.13
58.40
57.88
58.63
58.16
57.63
58.32
57.71
57.33
58.06
6.01
5.57
5.84
5.23
6.13
6.60
5.63
6.29
5.90
5.49
Fig. 6. Dataflow for a cross-validation run. Rules are generated from
InterPro families in SWISS-PROT having trailing accession number
digits 0–8, their confidence is tested against a given threshold (90
and 67%) and they are virtually applied on proteins in SWISS-PROT
having trailing accession number digit 9.
948
279 926
172 552
10 126
58.02
5.87
Protein fragments in the training set also reduce the
data quality. Suppose having a number of proteins with
a common sequence signature that induces a certain
annotation. If there are only sequence fragments of those
proteins contained in the database some might show the
pattern, others might not depending on which part of the
sequence is covered by the fragment. This is a highly
random process introducing noise into the training set
and their removal leads to lower error rates at the cross-
validation. But it also leads to a bias between training
data and target data, since the latter obviously will contain
fragments as well as whole protein sequences.
Cross-validation has been performed on both, training
sets including and excluding fragments. This was done by
splitting the whole set of proteins contained in SWISS-
PROT into ten parts of almost equal size. SWISS-PROT
accession numbers always end with a digit. The digit does
not encode any information about the protein as such and
was therefore used as the split criterion. Nine parts of the
split were used as training set to generate the decision
trees, which were tested on the remaining tenth part called
the test set. This procedure was repeated ten times, each
time changing training and test sets. In each run the
coverage and the error rate of the generated decision trees
was measured. As stated above, not all rules are useful
to be applied on unknown data. Rules can be selected
using the confidence criterion described in Formula 1 to
increase the reliability by decreasing the coverage and
vice versa. Two tests were performed to test the value of
that criterion and its influence on the observed error rate
in the cross-validation: the first test used rules having a
confidence of over 90% and the second test used rules with
a confidence of over 67%. The numerical results are shown
in Tables 1–4.
DISCUSSION
Reliability of the results
Obviously, the values of the observed error rate using
the cross-validation give better results than the confi-
dence obtained from Formula 1 would suggest. This
is due to the very careful calculation of this criterion
(z = 1.96) and could indicate a non-statistical distribution
of SWISS-PROT proteins allowing the supposition of
924

Automatic rule generation for protein annotation
Table 3. Fragments excluded in trainings set, confidence > 90%
44.9%
Insecta
1.4%
TrEMBL
SWISS– PROT
9.4%
Viridiplantae
End
digit
No. of
keywords
Covered
keywords
No. of
errors
% covered
% errors
0.0%
23.4%
Vertebrata
Rodentia
Primates
91.6%
0
1
2
3
4
5
6
7
8
9
25 471
25 303
25 275
25 418
24 824
25 338
25 676
25 310
25 158
25 590
8 443
8 381
8 597
8 480
8 286
8 542
8 600
8 277
8 335
8 348
121
120
94
102
138
155
101
146
123
136
32.67
32.65
32.64
32.96
32.82
33.10
33.10
32.13
32.64
32.09
1.43
1.43
1.09
1.20
1.67
1.81
1.17
1.76
1.48
1.63
4.7%
10.3%
17.8%
28.2%
39.4%
Mammalia
84.5%
0
20
40
60
80
100
Fig. 7. Distribution of the Taxa from which the proteins in
253 363
84 289
1236
32.78
1.47
IPR000301 descend (selection).
Table 4. Fragments excluded in trainings set, confidence > 67%
25.0%
Pfam PF00039
95.4%
75.0%
PRINTS PR00012
18.2%
End
digit
No. of
keywords
Covered
keywords
No. of
errors
% covered
% errors
32.1%
Prosite PS01253
100.0%
0
1
2
3
4
5
6
7
8
9
25 471
25 303
25 275
25 418
24 824
25 338
25 676
25 310
25 158
25 590
15 632
15 578
15 482
15 679
15 334
15 560
15 773
15 529
15 285
15 671
803
752
762
668
808
907
769
845
775
748
58.22
58.59
58.24
59.06
58.52
57.83
58.44
58.02
57.68
58.32
5.14
4.83
4.92
4.26
5.27
5.83
4.88
5.44
5.07
4.77
10.7%
PF00039+PR00012
18.2%
21.4%
PF00039+PS01253
95.4%
10.7%
PR00012+PS01253
18.2%
TrEMBL
10.7%
18.2%
SWISS–PROT
all
0
20 40
60
80
100
253 363
155 523
7837
58.29
5.04
Fig. 8. Distribution of protein signatures in IPR000301 (selection).
higher confidences in rules than the statistics assuming
a random distribution would suggest, e.g. finding four
proteins matching to a common signature and sharing
the same annotation without finding a counter-example
produces on the average a rule with an error rate much
less than the predicted 51% from Formula 1.
Furthermore, for the cross-validation it was assumed
that the information in SWISS-PROT is true and with-
out errors. Clearly, this precondition is not completely
fulfilled, leading to an increased error rate for the cross-
validation. In fact, there are three possible error sources
that contribute to inconsistencies in the cross-validation:
Points 2 and 3 suggest that the real error rate is less than
the one observed in the cross-validation.
The method assumes equal distribution of the proteins
in the training set (SWISS-PROT proteins in an InterPro
entry) and the data to be classified (TrEMBL entries or yet
unknown sequences matching the same InterPro entry).
This is clearly not the case as Figures 7 and 8 indicate.
This distribution is a result of the fact that TrEMBL
proteins are not randomly chosen to be annotated and
transferred to SWISS-PROT. For example, there is a high
interest in human proteins, which leads to their over-
representation in comparison to all other species.
Often whole protein families are annotated or updated.
For some families, all proteins share the same signatures,
which leads to an over-representation of these signatures
in SWISS-PROT. Generating rules from these sets and
applying them on proteins of different origin or matching
different additional patterns might lead to systematic
errors. An improved method of validation based on
(1) The rule suggests a Keyword which is not contained
in a target due to a biological reason (True error).
(2) The Keyword has been forgotten to be annotated
(Inconsistency in SWISS-PROT).
(3) The protein does not match the precondition of the
rule due to a FP match to one of the signature
databases (Inconsistency in SWISS-PROT).
925

E.Kretschmann et al.
empirical evidence is under construction and is planned
to be implemented in later versions of this tool.
At the current status, the rules can be used to support
the manual annotation process performed by the SWISS-
PROT database curators. It is also ready to be made publi-
cally available (http://golgi.ebi.ac.uk:8080/Spearmint/).
Further developments
There are two ways for further developments. Firstly, the
rule generation process can be improved and secondly, the
rules can be applied in different applications. Currently,
the following projects are under development:
REFERENCES
Apweiler,R. (2000) Protein sequence databases. Adv. Protein
Chem., 54, 31–71.
Apweiler,R. (2001) Functional information in SWISS-PROT: the
basis for large-scale characterisation of protein sequences. Brief-
ings in Bioinformatics, 2, 9–18.
• A web-based application that allows application of the
rules not only on TrEMBL entries, but also on raw
amino acid sequences.
Apweiler,R., Attwood,T.K., Bairoch,A., Bateman,A., Birney,E.,
Biswas,M., Bucher,P., Cerutti,L., Corpet,F., Croning,M.D.R,
Durbin,R., Falquet,L., Fleischmann,W., Gouzy,J., Herm-
jakob,H., Hulo,N., Jonassen,I., Kahn,D., Kanapin,A., Kar-
avidopoulou,Y., Lopez,R., Marx,B., Mulder,N.J., Oinn,T.M.,
Pagni,M., Servant,F., Sigrist,C.J.A. and Zdobnov,E. (2000)
InterPro—an integrated documentation resource for protein
families, domains and functional sites. Bioinformatics, 16,
1145–1150.
• Mining for Description-, Comment-, and Feature
Lines.
• Integration of other data mining techniques.
• Application of rules on entries in Ensembl (http://
www.ensembl.org/) to predict functionality.
Attwood,T.K.,
Croning,M.D.R.,
Flower,D.R.,
Lewis,A.P.,
Mabey,J.E., Scordis,P., Selley,J.N. and Wright,W. (2000)
PRINTS-S: the database formery known as PRINTS. Nucleic
Acids Res., 28, 225–227.
• Automated application of the rules on proteins in
TrEMBL without human interaction.
Bairoch,A. and Apweiler,R. (2000) The SWISS-PROT protein
sequence database and its supplement TrEMBL in 2000. Nucleic
Acids Res., 28, 45–48.
Bateman,A., Birney,E., Durbin,R., Eddy,S.R., Howe,K.L. and
Sonnhammer,E.L.L. (2000) The Pfam protein falilies database.
Nucleic Acids Res., 28, 263–266.
Corpet,F., Servant,F., Gouzy,J. and Kahn,D. (2000) ProDom and
ProDom-CG: tools for protein domain analysis and whole
genome comparisons. Nucleic Acids Res., 28, 267–269.
Fleischmann,W., Mo¨ller,S., Gateau,A. and Apweiler,R. (1999) A
novel method for automatic functional annotation of proreins.
Bioinformatics, 15, 228–233.
The Gene Ontology Consortiuum (2000) Gene Ontology: tool for
the unification of biology. Nature Genet., 25, 25–29.
Hofmann,K., Bucher,P., Falquet,L. and Bairoch,A. (1999) The
PROSITE database, its status in 1999. Nucleic Acids Res., 27,
215–219.
• Automated rule generation on GO terms (The
Gene Ontology Consortiuum, 2000) rather than on
Keywords.
One of the most imperative tasks is to achieve an improved
confidence calculation. The current routine does not use
information about the number of TNs or FNs in the
calculation and hence prefers the generation of frequent
keywords rather than rare ones, i.e. general Keywords are
produced more often than specific ones, but the latter are
the more valuable ones. Extracting reliable rules for rare
Keywords will certainly be a very useful improvement of
the tool.
For all calculations independence between training and
target set were assumed, which is clearly not the case. An
empirical test to collect data about the influence of the bias
is helpful to get a better picture about the performance of
the rules on unknown data.
Hyafil,L. and Rivest,R.L. (1976) Constructing optimal binary deci-
sion trees is NP-complete. Inf. Process. Lett. 5, 1, 15–17.
Junker,V., Apweiler,R. and Bairoch,A. (1999) Representation of
functional information in the SWISS-PROT data bank. Bioinfor-
matics, 15, 1066–1067.
CONCLUSION
The presented method mines for Keyword annotation in
SWISS-PROT using a Java implementation of the C4.5
algorithm on protein groups assembled in InterPro entries.
The results are satisfactory in terms of coverage and
confidence, yet it was pointed out that both aspects can
be further improved. Including other methods to group
proteins into sets containing similar proteins like CluSTr,
Prodom (Corpet et al., 2000) or others will help to
increase the coverage while a refined statistical analysis
will improve ordering of the generated rules in terms of
reliability. This is supposed to lead to higher values of
confidence.
Krause,A., Nicode`me,E., Bornberg-Bauer,M., Rehmsmeier,M. and
Vingron,M. (1999) WWW access to the SYSTERS protein
sequence cluster set. Bioinformatics, 15, 262–263.
Kriventseva,E.V., Fleischmann,W., Zdobnov,E.M. and Apweiler,R.
(2001) CluSTr: a database of clusters of SWISS-PROT +
TrEMBL proteins. Nucleic Acids Res., 29, 33–36.
Mo¨ller,S., Leser,U., Fleischmann,W. and Apweiler,R. (1999)
EDITtoTrEMBL: a distributed approach to high-quality auto-
mated protein sequence annotation. Bioinformatics, 15, 219–227.
Quinlan,,J.R. (1986) Induction of decision trees. Mach. Learn., 1,
81–106.
Quinlan,J.R. (1993) C4.5: Programs for Machine Learning. Morgan
Kaufmann, San Francisco, CA.
926