Evaluation of N-myristoylation Prediction Tool using Machine Learning

Abstract

Protein sequences constitute molecular complex in an organism. However it is difficult to find a sequence rule such as cascade reaction signals, post translational modi-fication signals and so on. These sequence signals perform an essential role in regulating cellular structure and func-tion. In previous study, we could find sequence rules of N-myristoylated proteins easily with computational approach. Subsequently, we have developed a CGI tool to predict N-myristoylated proteins with their sequence rules. In this study, we performed accuracy evaluation of our developed CGI tool. As a result, we show that developed CGI tool predict N-myristoylated proteins effectively with characteristics of N-myristoylated protein sequences.

Full text

Page 1

Evaluation of N-myristoylation Prediction Tool using Machine Learning
Sayaka Kado†　 Ryo Okada††　 Manabu Sugii†††　 Hiroshi Matsuno†　 Satoru Miyano††††
†Graduate School of Science and Engineering, Yamaguchi University, Japan
††Network Solution Group, Hitachi Chugoku Solutions, Japan
†††Madia and Information Technology Center, Yamaguchi Unibersity, Japan
††††Human Genome Center, University of Tokyo, Japan
†1677-1 Yoshida, Yamaguchi 753-8513, Japan
††11-10, Motomachi, Hiroshima 730-0011, Japan
†††1677-1 Yoshida, Yamaguchi 753-8513, Japan
††††4-6-1 Shirokanedai, Minatoku, Tokyo 108-8639, Japan
†Tel/Fax: +81-83-933-5697
E-mail: manabu@yamaguchi-u.ac.jp
Abstract: Protein sequences constitute molecular complex
in an organism. However it is difficult to find a sequence
rule such as cascade reaction signals, post translational modi-
fication signals and so on. These sequence signals perform
an essential role in regulating cellular structure and func-
tion. In previous study, we could find sequence rules of N-
myristoylated proteins easily with computational approach.
Subsequently, we have developed a CGI tool to predict N-
myristoylatedproteinswiththeirsequencerules. Inthisstudy,
we performed accuracy evaluation of our developed CGI tool.
As a result, we show that developed CGI tool predict N-
myristoylated proteins effectively with characteristics of N-
myristoylated protein sequences.
Introduction
Protein N-myristoylation is a lipid modification of pro-
teins. Figure 1 shows a reaction process of protein N-
myristoylation. Protein N-myristoylation is a cotranslational
protein modification catalyzed by two enzymes, methionime
aminopeptidase and N-myristoyltransferase(NMT). The ini-
tial methionine(Met) is cleaved from N-terminus of amino
acid sequences by a methionine aminopeptidase, and then the
myristic acid is linked to exposed glycine via an amide bond
by NMT. NMT catalyzes the transfer of myristic acid from
myristoyl-CoA to the N-terminus glycine residue of substrate
protein. Most of myristoylated proteins have a physiologi-
cal activity such as cell signaling protein, expressing specific
functions through binding organelle membrane. It is known
that membrane binding reaction mediated by myristoylation
is controlled variedly, and play a crucial role in functional
regulation mechanisms of proteins in cell signaling pathway.
N-myristoylated proteins have a specific sequence at the N-
terminus called N-myristoylation signal sequence, and this
sequence is probably composed of 6 to 9 amino acids (up to
17)[1]. However, itisveryhardtopredicttheN-myristoylated
proteins with the signal sequence among the genomic pro-
tein database, because the information on the amino acid se-
quences is very vast, and N-myristoylation is not based on
one simple rule but many specific rules. We had applied the
machine learning system BONSAI to predict patterns based
on positive and negative examples to be classified. Figure 2
shows the mechanism of machine learning system BONSAI.
Furthermore, Okada et al.[2] developed a CGI tool that pre-
dicts whether a given protein is N-myristoylated or not.
Figure 1: Reaction process of protein N-myristoylation.
The 23rd International Technical Conference on Circuits/Systems,
Computers and Communications (ITC-CSCC 2008)
133

Page 2

In order to examine the performance of our CGI tool, we
performed an experimentation using sample data of amino
acidsequences thathad beenconfirmedto beN-myristoylated
or not by biochemical approach. Our CGI tool could iden-
tify N-myristoylated protein using predicted features of N-
myristoylation signals and could classify sequences into pos-
itive and negative categories rapidly and correctly.
Figure 2: Machine learning system BONSAI.
A Prediction Tool
with Computational Approach
Computational techniques are useful for predicting rules
from a huge amount of data on the sequence required for
N-myristoylation. The machine learning system BONSAI,
which finds rules in the combination of alphabet indexings
and decision trees, is applied to this CGI tool. BONSAI
is a system for knowledge acquisition based on the theory
of Probably Approximately Correct Learning (PAC learn-
ability) and uses the method of local search[3]. The alpha-
bet indexing groups letters in positive and negative example
by mapping these letters to fewer numbers of letters. This
CGI tool predicted N-myristoylated proteins with a focus
on the N-myristoylation signal sequence important to be N-
myristoylated. This CGI tool is composed with 4 decision
trees that were provided with positive and negative examples
by BONSAI and has high accuracy rate of the classification.
Figure 3 shows one of 4 decision trees used our developed
CGI tool. The positive examples include 78 myristoylated
amino acid sequences, and the negative examples include
800 amino acid sequences randomly selected from the hu-
man protein in NCBI GenBank database[4]. The reason that
we used randomly selected protein sequences as negative ex-
ample is that no protein sequence being not N-myristoylated
protein has been known. As a result, the CGI tool could
identify N-myristoylated protein using predicted features of
N-myristoylation signals and could classify sequences under
two categories rapidly and correctly.
Figure 4 shows a system flowchart of our developed CGI
tool. UsersinputthedataofN-terminusaminoacidsequences
to our tool. And, users select one or more decision trees for
the prediction. Our CGI tool removes the initial metchionine
and glycine from the N-terminus of given sequences because
N-myristoylated proteins surely have these two amino acids
at the N-terminus. Subsequently, our CGI tool convert alpha-
bets of amino acid sequences into numbers based on alphabet
indexings(indexing, in short), provided at each decision tree .
Alphabet indexing is the transformation of symbols to reduce
the number of letters assigned to positive and negative exam-
ples without omitting important information in the original
data in order to find the pattern of N-myristoylation signals.
In the case of amino acid residues, alphabet indexing can be
regarded as a classification of 20 kinds of amino acid residues
to a few categories. Indexing contributes not only quicken the
computations involved in finding rules but also to simplify
expression patterns assigned at the nodes of decision trees.
After that, our CGI tool classifies proteins into N-
myristoylated proteins and others according to rules of deci-
sion trees with indexed amino acid sequences. Finally, users
obtain prediction results that gives the classification. Our
CGI tool can predict the N-myristoylation without a machine
learning process because we provided decision trees before-
hand, which can classify the N-myrisoylated protein more ex-
actly and adapted the these decision trees to our CGI tool.
This contributes to reduce the time for processing the given
inputs.
Figure 3: One of 4 decision trees used our CGI tool.
134

Page 3

Figure 4: System flowchart of our CGI tool.
Experiment
In this study, we performed accuracy evaluation of the de-
veloped CGI tool. We used amino acid sequences that are
confirmed to induce N-myristoylation biologically, and con-
ductedaccuracycomparisonwithotherpredictedtoolpredict-
ing protein N-myristoylation called ”NMT-The predictor”[5].
NMT-The predictor classifies proteins into N-myristoylated
and others searching a motif of protein N-myristoylation with
analysis of substrate sequences and kinetic data of predicted
proteins. We focused on the results of two tool’s predictions
and the features of the pattern whose sample sequences were
classified into positive and negative categories using decision
trees. We used 78 protein amino acid sequences of eukary-
ota and 44 protein amino acid sequences of A.Thaliana that
have been confirmed to be induced N-myristoylation biolog-
ically, and 88 protein amino acid sequences of TNF(Tumor
necrosis factor) which have been confirmed not to be induced
N-myristoylation biologically.
Additionally, our developed CGI tool and NMT-The pre-
dictor need to configure some parameters to predict. On our
developed tool, users need to make one or more selections of
using decision trees for the prediction. It is possible for users
to configure 15 paremeters in all. On the other hand, users
need to select one of 2 parameters to identify species of or-
ganisms, eukaryote parameter and fungi parameter, on NMT-
The predictor. Accordingly, we try to evaluate an accuracy
rate between two systems with all of 17 parameters.
We classified prediction results into 3 categories. ”Induce
N-myristoylation” is a rate of all protein sequences that were
predicted to induce N-myristoylation. ”Can not classify N-
myristoylated or not” is a rate of all protein sequences that
could not be classified a N-myristoylated proteins and oth-
ers clearly. ”Non-induce N-myristoylation” is a rate of all
protein sequences that were predicted not to be induced N-
myristoylation.
Result and Discussion
Our N-myristoylation prediction tool showed high accu-
racy rate of classification for N-myristoylation and almost the
same accuracy rate of classification for most of the 210 pro-
tein sequences in this evaluation experiment (Fig.7, 8, 9).
Figure 5: Prediction result of 78 myristoylated amino acid
sequences of eukaryote.
Figure 6: Prediction result of 88 not myristoylated amino acid
sequences of TNF.
And, we showed more a beneficial result of our tool than
that of NMT-The predictor. NMT-The predictor could not
obtain high accuracy rate of classification using 44 protein
amino acid sequences of A.Thaliana. We considered that
the parameter setting of NMT-The predictor is not applied to
135

Page 4

the prediction of protein N-myristoylation on plant species.
On the other hand, our CGI tool provided the high accuracy
rate of classification using all sequence samples (Fig.9). Our
CGI tool used positive and negative examples of protein se-
quences of several species for the machine learning process in
which the decision trees developed by Okada et al. are used.
Therefore, our CGI tool could identify common rules for N-
myristoylation that do not depend on species of organisms.
Figure 7: Prediction result of 44 myristoylated amino acid
sequences of A.Thaliana.
In addition, we investigated what sequence was recognized
by decision trees adopted our CGI tool in order to clas-
sify the protein sequence. Our CGI tool could predict N-
myristoylation signal sequence according to features of the
amino acid at position 6 from the N-terminus that is impor-
tant position for N-myristoylation. Consequently, we suggest
that our CGI tool can predict N-myristoylated proteins effec-
tively using characteristics of amino acid sequences that are
important for protein N-myristoylation.
We considered that NMT-The predictor have been devel-
oped base on N-myristoylation signal sequences as its own
database, and can predict N-myristoylation signal sequences
correctly if the database has the sequence information about
species of oranisms of query. However our developed tool
canpredictN-myristoylationsignaleffectivelywithoutthese-
quence information, our developed tool has the advantage of
classfying unknown protein correctly.
References
[1] Maurer-Stroh, S., Eisenhaber, B., and Eisenhaber, F.,
“N-terminal N-myristoylation of proteins: prediction of
substrate proteins from amino acid sequence”, J. Mol.
Biol., 317(4):541-557, 2002
[2] Sugii，M., Okada, R., Matsuno, H., Mayano, S., “Per-
formance Improvement in Protein N-Myristoyl Classi-
fication by BONSAI with Insignificant Indexing Sym-
bol”, Genome Inform., Vol.32, pp.277-286, 2007
[3] Shimozono, S., Shinohara, A., Shinohara, T., Miyano,
S., Kuhara, S., and Arikawa, S., “Knowledge Acquisi-
tion from Amino Acid Sequences by Machine Learning
System BONSAI”, Trans. Inform. Process. Soc. Japan,
Vol.35, pp.2009-2018
[4] NCBI : ftp://ftp.ncbi.nih.gov/
[5] NMT : http://mendel.imp.ac.at/myristate
136