Predicting Moonlighting Proteins from Protein Sequence
Jing Hu
*
and Yihang Du
Department of Computer Science, Franklin & Marshall College, Lancaster, PA, U.S.A.
Keywords: K-Nearest Neighbors Method, Bit-Score Weighted Euclidean Distance, Feature Selection, Moonlighting
Proteins.
Abstract: High-throughput proteomics projects have resulted in a rapid accumulation of protein sequences in public
databases. For the majority of these proteins, limited functional information has been known so far.
Moonlighting proteins (MPs) are a class of proteins which perform at least two physiologically relevant
distinct biochemical or biophysical functions. These proteins play important functional roles in enzymatic
catalysis process, signal transduction, cellular regulation, and biological pathways. However, it has been
proven to be difficult, time-consuming, and expensive to identify MPs experimentally. Therefore,
computational approaches which can predict MPs are needed. In this study, we present MPKNN, a K-nearest
neighbors method which can identify MPs with high efficiency and accuracy. The method is based on the bit-
score weighted Euclidean distance, which is calculated from selected features derived from protein sequence.
On a benchmark dataset, our method achieved 83% overall accuracy, 0.64 MCC, 0.87 F-measure, and 0.86
AUC.
1 INTRODUCTION
Due to high throughput sequencing technologies,
huge number of protein sequences have been
accumulated in public databases, waiting to be
analyzed. The study of protein functions is important
for the understanding of the cellular mechanism and
biological pathways of organisms. Most proteins as
we know only exhibit single function. However, there
are a class of proteins named moonlighting proteins
(MPs) in which a single protein performs multiple
physiologically relevant distinct biochemical or
biophysical functions that are not result of gene
fusions, alternative RNA splicing, multiple domains,
DNA rearrangement, or proteolytic fragments
(Jeffery, 2015; Jeffery, 2018; Weaver, 1998; Jain,
2018; Shirafkan, 2021). Since the first MP was
detected in 1980s (Piatigorsky, 1989), scientists have
found MPs in many types of species, including
bacteria, archaea, mammals, reptiles, birds, fish,
worms, insects, plants, fungi, protozoans and even
viruses (Jeffery, 2018). The datasets of these proteins
are stored in public databases such as MoonProt
(Chen, 2018), MultitaskProtDB-II (Franco-Serrano,
2018), and MoonDB (Ribeiro, 2019). Recent studies
*
https://www.fandm.edu/jing-hu
have found that a large number of MPs play important
roles in diseases, infection, virulence, or immune
responses, and some can be potential vaccination
targets (Jeffery, 2015). Therefore, identification and
study of MPs is crucial for research on diseases and
drug-target discovery.
However, it is time-consuming and expensive to
identify MPs experimentally in laboratory. Therefore,
computational methods which can predict MPs with
high performance are in urgent need. As to our
knowledge, there are only a limited number of
methods published so far. Chapple et al. has
developed MoonGO (Chapple, 2015), a method
based on features derived from protein-protein
interaction networks and gene ontology (GO)
information to predict MPs. MPFit (Khan, 2016) is a
machine learning method which utilizes features
derived from gene ontology (GO), protein–protein
interactions, gene expression, phylogenetic profiles,
genetic interactions and network-based graph
properties to protein structural properties to predict
MPs. Although they reported a high accuracy of 98%
in identifying MPs when GO information was
incorporated, the prediction performance was
degraded dramatically to 75% accuracy when GO
270
Hu, J. and Du, Y.
Predicting Moonlighting Proteins from Protein Sequence.
DOI: 10.5220/0011782300003414
In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 3: BIOINFORMATICS, pages 270-275
ISBN: 978-989-758-631-6; ISSN: 2184-4305
Copyright
c
2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
annotations were removed. Jain et al. proposed a text
mining method called DextMP to predict MPs from
several different information sources, database
entries, literature, and large-scale omics data (Jain,
2018). However, the lack of GO annotations and
literature-based information for the majority of
proteins restricted the application of the methods
mentioned above. Recently, Shirafkan et al.
attempted to predict MPs using features extracted
only from protein sequences (Shirafkan, 2021). After
assessing 8 different machine learning methods with
each of 37 distinct feature vectors to detect
moonlighting proteins, they found that the Support
Vector Machine (SVM) model based on the split
amino acid composition (SAAC) feature set has the
highest accuracy of 77% in classifying between MPs
and non-MPs
In this paper, we propose MPKNN, a K-nearest
neighbors (KNN) method to predict MPs. Our
method relies on the bit-score weighted Euclidean
distance, which is calculated from selected features
including compositions of certain amino acids,
extended pseudo-amino acids (Chou, 2003) (Du,
2006), and FEGS features (Mu, 2021). The final
method achieved 83% overall accuracy, 0.64 MCC,
0.87 F-measure, and 0.86 AUC on a benchmark
dataset used in previous study (Shirafkan, 2021).
2 METHODS AND MATERIAL
2.1 Dataset
We used the benchmark dataset that was used in the
previous study (Shirafkan, 2021). In their study, all
moonlighting proteins were experimentally verified
and collected from MoonProt (Chen, 2018) database.
After removal of sequence redundancy using CD-hit
(Fu, 2012) with a 40% mutual sequence similarity
threshold, there were 315 proteins left in the final
dataset, among which 215 were moonlighting
proteins and 136 were non-moonlighting proteins
from species including Mus Musclus, Human, E. coli,
Yeast, Rat, Drome, Arath, and others.
2.2 Feature Extraction
We have investigated feature sets such as amino acid
composition, extended pseudo-amino acid
composition (Chou, 2003) (Du, 2006), and FEGS
features (Mu, 2021) in this study.
2.2.1 Amino Acid Composition
The amino acid composition of a protein sequence
was calculated by
=
(1)
where
and
are the numbers of amino acid i and
j in the protein sequence.
2.2.2 Extended Pseudo-Amino Acid
Composition
The original model of pseudo-amino acid
composition (Chou, 2003) consists of compositions
of 20 amino acids in a protein and λ different ranks of
sequence-order correlation factors (i.e., delta function
set). In this study, we extended the definition of
pseudo-amino acid composition by including 9 more
sets of various physicochemical properties that were
investigated in the previous study of Du and Li (Du,
2006).
In this study, the delta function set was calculated
as in (Chou, 2003). Suppose a protein with a
sequence of L amino acid residues:
…
…
,
where
represents the amino acid at sequence
position i. The first set, delta-function set, consisted
of λ sequence-order-correlated factors, which were
given by
=
1
−
Δ
,
(2)
where =1,2,3…,< and Δ
,
=
Δ
,
=1 if
=
, 0 otherwise. These
features were named as {
,
,…,
}.
The remaining 9 sets of physicochemical
properties were based on AAindex values
(Kawashima, 2000). Similar as (Du, 2006), the
following AAindex indices were used: BULH740101
(transfer free energy to surface), EISD840101
(consensus normalized hydrophobicity),
HOPT810101 (hydrophilicity value), RADA880108
(mean polarity), ZIMJ680104 (isoelectric point),
MCMT640101 (refractivity), BHAR880101 (average
flexibility indices), CHOC750101 (average volume
of buried residue), COSI940101 (electron-ion
interaction potential values). For each of 9 AAindex
indices, we obtained
μ
sequence-order-correlated
factors by
ℎ
=
1
−
,
(3)
Predicting Moonlighting Proteins from Protein Sequence
271
where =1,2,3,…,,<, and
,
=(
)∙
(
). In this study, (
) and (
) are the
normalized AAindex values of residues R
i
and R
j
respectively. The normalized AAindex value of each
amino acid was calculated by applying
(
)
=
(
)
−
∑
−
20
(4)
where =1,2,3,...20.
(
) is the original
AAindex value of amino acid i, and
is the average
AAindex value of 20 amino acids. For each of 9
AAIndex types (i.e., BULH740101, EISD840101,
etc.), we obtained μ features using (3) and (4). In total
there are 9 features. We named these features as
{BULH740101_1, BULH740101_2, …
BULH740101_μ, EISD840101_1, EISD840101_2,
… EISD840101_μ, … COSI940101_1,
COSI940101_2, … COSI940101_μ}. Therefore, the
pseudo-amino acid compositions consist of λ (delta-
function factors) + 9 (9 sets of physicochemical
factors) numbers. In this study, both λ and
μ
were set
to 10.
2.2.3 FEGS Features
FEGS (Feature Extraction based on Graphical and
Statistical features) is a protein feature extraction
model by integrating the graphical representation of
protein sequences based on the physicochemical
properties of amino acids and statistical features of
the protein sequences (Mu, 2021). Using the
MATLAB script provided by the authors, a set of 578
features was extracted for each protein sequence in
our dataset.
In total, we have investigated 698 sequence-
derived features (i.e., 20 amino acid compositions, 10
delta-function factors, 9 sets of physicochemical
factors with 10 features in each set, and 578 FEGS
features).
2.3 Bit-Score Weighted Euclidean
Distance
For each query protein t, its distance to a training
protein T is calculated as
,
=
(
−
)
(,)
(5)
where t
i
and T
i
are the
i
th
features (i.e., amino acid
composition, pseudo-amino acid composition, FEGS
features) of the query protein t and the training protein
T respectively, and N is the number of features used.
BS(t, T) is the bit score computed by the blastp
program of Blast package (Altschul, 1997) when
comparing the local sequence similarity between
protein sequences t and T. It is a normalized score
which is independent of query sequence length and
database. A higher bit score indicates two protein
sequences are more similar, and vice versa. Notice
that
∑(
−
)
gives the Euclidean distance
between two proteins. Here, the distance is weighted
by a factor (i.e., bit score). Therefore, the distance is
referred to as the bit-score weighted Euclidean
distance (BS-WED).
As can be seen from the
equation (5), the more similar the corresponding
features, the higher the sequence similarity between
two proteins, the smaller the distance (as measured by
BS-WED).
2.4 K-Nearest Neighbors Method
A traditional K-nearest neighbors method (KNN)
method classifies the query sample by the majority
voting strategy. For each query sample, KNN finds its
k nearest neighbors in the training dataset, and then
assigns it to the class to which most of its neighbors
belong. The KNN method used in this study differs in
that it finds k nearest neighbors to the query sample
from each class of training samples. The average of
these k distances is calculated as the adjacency value
of the query sample to that class. The adjacency
values between the query sample and all classes are
compared and the query sample is assigned to the
class to which it has the smallest adjacency value.
Specifically, for each test protein, its distance
(measured by BS-WED) to every moonlighting
protein in the training set was calculated using (5).
The k shortest BS-WEDs were chosen, namely d
mp-1
,
d
mp-2
,…, d
mp-k
. Then, the adjacency value between the
test protein and the MP class (denoted as
) was
given by
=
/
(6)
The adjacency value between the query protein
and non-MP class (
) was calculated in a
similar way. Then the query protein was predicted as
moonlighting protein if
<
⁄
; non-
moonlighting protein otherwise. The default value of
θ was set to 1.
BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms
272
2.5 Performance Measurement
Ten-fold cross-validation was used to evaluate the
performance of the algorithm. In this study, MPs were
treated as the positive samples, while non-MPs were
treated as the negative samples.
Performance was measured using recall,
precision, Acc (overall accuracy), MCC (Matthews
Correlation Coefficient), F-Measure, and AUC (area
under the ROC curve). F-measure and MCC are
balanced measurements of the performance on both
positive and negative samples. AUC is the area under
the ROC curve when adjusting true positive rate vs.
false positive rate by tuning the parameter θ.
= (+)
⁄
(7)
= (+)
⁄
(8)
=
+
+++
(9)
−=
2××
+
(10)
=
×−×
(+)(+)(+)(+)
(11)
2.6 Heuristic Feature Selection Process
For each protein, there were 20 (classic amino acid
compositions) + λ (delta-function factors) + 9
μ
(9 sets
of physicochemical factors) + 578 (FEGS) extracted
features. In this study, λ and
μ
were both set to 10.
Therefore, there were 698 features investigated in
total. To identify the most useful subset of features, a
previously described greedy feature selection
algorithm by our group (Hu, 2012) was employed.
The greedy search was based on the wrapper
approach and it started with a feature set that included
20 amino acids. Let n be the size of the feature set.
Then =20at the beginning. The algorithm could
be divided into two stages: reduction and growth. In
the reduction stage, the size of the feature set was
gradually reduced. First, one amino acid was removed
and the composition of the remaining n-1 amino acids
were used to calculate WEDs. Ten-fold cross-
validation was used to evaluate the performance of
the method. This step was repeated n times, so that
every combination of n-1 amino acids was tried. The
combination that improved the performance most was
chosen. Thus, the size of the feature set was reduced
from n to n-1. This reduction process was continued
until removing any amino acid from the feature set
would reduce the performance. At the end of the
reduction stage, we reached a feature set that included
the composition of N amino acids (≤20 ). Then,
we used a growth stage to increase the size of the
feature set by adding pseudo-amino acid
compositions. One pseudo-amino acid composition
was added at a time and the resulting feature set was
used to calculate WEDs. Ten-fold cross-validation
was used to evaluate the performance of the method.
The pseudo-amino acid composition that brought the
biggest improvement in performance was chosen and
added into the feature set. Thus, the size of the feature
set was increased to N+1. This growth process
continued until adding any more pseudo-amino acid
composition would decrease the performance. Then
we began the process of adding FEGS feature one at
a time until adding any more FEGS feature into the
set would decrease the prediction performance. In the
end, we obtained a set of 18 features that include
compositions of 12 amino acids {A, N, D, C, E, G, I,
L, K, F, S, T}, 4 delta-function factors {
,
,
,
},
and 2 physicochemical factors {EISD840101_2 and
HOPT810101_10}.
3 RESULTS
The proposed MPKNN method was first used to
classify between MPs and non-MPs using only the
compositions of 20 amino acids. We then assessed the
prediction ability of each feature set individually and
also that of all combined features. Finally, we seek to
improve the prediction performance by applying the
heuristic feature selection process mentioned above
to choose the more relevant features. Our final
method was also compared with previously published
method (Shirafkan, 2021) on the same benchmark
dataset.
Table 1: Prediction performance of each feature set.
Methods Acc MCC
Amino acid compositions 80.3% 0.579
delta-function factors 73.8% 0.432
BULH740101 76.4% 0.490
EISD840101 80.1% 0.572
HOPT810101 77.5% 0.522
RADA880108 77.8% 0.523
ZIMJ680104 73.5% 0.429
MCMT640101 72.1% 0.397
BHAR880101 70.7% 0.360
CHOC750101 72.6% 0.412
COSI940101 71.2% 0.374
EGS features 79.5% 0.560
All features combined 78.3% 0.535
Predicting Moonlighting Proteins from Protein Sequence
273
3.1 Predicting MPS Using Only Amino
Acid Composition
First only compositions of 20 amino acids were used
to calculate the distance (i.e., BS-WED) of each test
protein to each training protein. Ten-fold cross-
validation was used to evaluate the performance on
the dataset. The parameter θ was set to 1. Various k
values ranging from 1 to 30 were tried. For
comparison, we also searched for the best
performance of traditional KNN method using the
standard Euclidean distance for various k values in
the same range (i.e., 1-30). As can be seen from Table
1 (row 2), MPKNN achieved the performance of
80.3% accuracy and 0.579 MCC when k = 20. In
comparison, standard KNN has the most optimal
performance of 76.4% accuracy and 0.490 MCC
when k = 19. Therefore, it is clear that BS-WED is a
better distance measurement than standard Euclidean
distance when measuring the relationship between the
query protein and the training proteins (i.e., MPs or
non-MPs). The detailed prediction performance of
MPKNN using only compositions of 20 amino acids
is shown in Table 2 (Column 2).
3.2 Prediction Performance of Each
Feature Set
The proposed KNN method based on BS-WED was
tested on each feature set individually to assess their
usefulness in predicting MPs. The parameter θ was
set to the default value (i.e., 1). Various k values
ranging from 1 to 30 were tried and the best
performance was kept. We also evaluated the
prediction performance of all features combined. The
results are listed in Table 1. It is obvious that the
prediction performance couldn’t be further improved
by simply combining all features investigated in this
study. This could due to the fact that not all features
were useful for the prediction. Also, some features
might be correlated with each other, which could
impair the prediction performance.
Table 2: Comparison of different methods.
Method MPKNN
+20 AAs
MPKNN
+ 18
selected
features
Shirafkan
et al.’s
method
Recall 91.6% 91.6% -
Precision 79.4%
82.4%
74%
Acc 80.3%
82.9%
77%
MCC 0.579 0.635 -
F-Measure 0.851 0.868 0.75
AUC 0.750
0.862
0.75
3.3 Performance After Feature
Selection
We then seek to improve the prediction performance
by applying the heuristic feature selection process as
described in Methods and Materials to search for a
subset of features that was (almost) most useful for
the prediction. In the end a set of 18 features that
include compositions of 12 amino acids {A, N, D, C,
E, G, I, L, K, F, S, T}, 4 delta-function factors
{
,
,
,
}, and 2 physicochemical factors
{EISD840101_2 and HOPT810101_10} were
chosen. Adding more features did not improve the
prediction performance. As can been seen from Table
2 (Column 3), using selected features, MPKNN
improved its performance to 91.6% recall, 82.4%
precision, 82.9% accuracy, 0.635 MCC, 0.868 F-
measure, and 0.862 AUC.
We also compared our final MPKNN (i.e., using
18 selected features) with previously published
method by (Shirafkan, 2021) on the same benchmark
dataset. The prediction performance of Shirafkan et
al.’s method was directly obtained from their report
(as shown in Table 2, Column 4). Table 2 clearly
shows that our method has achieved far more superior
performance than that of Shirafkan et al.
4 CONCLUSIONS
In this study we present MPKNN, a KNN method
which can predict MPs with 91.6% recall, 82.4%
precision, 82.9% accuracy, 0.635 MCC, 0.868 F-
measure, and 0.862 AUC. The method is based on a
bit-score weighted Euclidean distance (BS-WED) to
measure the similarity between proteins. Compared to
the standard Euclidean distance, BS-WED takes
account of both compositions and sequence
similarity.
The benchmark dataset used in this study was
relatively small, and therefore feature selection
(wrapper method) and cross-validation were
performed on the same dataset to avoid insufficient
training. To better estimate the generalization ability
of our method, it would be preferred to carry out these
two processes on two separate non-overlapping
datasets. For the future work, we plan to curate a more
comprehensive dataset with a larger number of MP
and non-MP proteins from various protein databases
so that we may split the dataset into two parts, with
the first part reserved for feature selection and the
second part for cross-validation (i.e., training and test)
using the selected features. We also plan to
investigate the possibility of improving the prediction
BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms
274
performance by including other features such as
evolutionary features (PSSM profiles.), predicted
structural information of each protein, feature vectors
calculated by the ftrCOOL library (Amerifar, 2020),
etc., and applying our method to identify potential
MPs in proteomes of human and other species.
ACKNOWLEDGEMENTS
This work was partially supported by the FR/PDF and
Hackman grant from Franklin & Marshall College.
REFERENCES
Altschul, S., et al. (1997). Gapped BLAST and PSI-
BLAST: a new generation of protein database search
programs. Nucleic Acids Res., vol. 25, pp. 3389-3402.
doi: 10.1093/nar/25.17.3389.
Amerifar, S. and Zahiri, J. (2020). ftrCOOL: Feature
extraction from biological sequences. CRAN - Package
ftrCOOL - R Project.
Chapple, C., et al. (2015). Extreme multifunctional proteins
identified from a human protein interaction network.
Nat Commun., 6:7412. doi: 10.1038/ncomms8412.
Chen, C., et al. (2018). MoonProt 2.0: an expansion and
update of the moonlighting proteins database. Nucleic
Acids Res., 46(D1): D640-D644. doi:
10.1093/nar/gkx1043.
Chou, K. and Cai, Y. (2003). Prediction and classification
of protein subcellular location-sequence-order effect
and pseudo amino acid composition. J. Cell Biochem.,
vol. 90, pp. 1250-1260. doi: 10.1002/jcb.10719.
Du, P. and Li, Y. (2006). Prediction of protein
submitochondria locations by hybridizing pseudo-
amino acid composition with various physicochemical
features of segmented sequence. BMC bioinformatics,
vol. 7 518. 30. doi:10.1186/1471-2105-7-518.
Franco-Serrano, L., et al. (2018) MultitaskProtDB-II: an
update of a database of multitasking/moonlighting
proteins. Nucleic Acids Res., 46(D1): D645-D648. doi:
10.1093/nar/gkx1066.
Fu, L., et al. (2012). CD-HIT: accelerated for clustering the
next generation sequencing data. Bioinformatics, 28
(23): 3150-3152. doi: 10.1093/bioinformatics/bts565.
Hu, J. and Ng, P. (2012). Predicting the effects of
frameshifting indels. Genome Biol., vol. 9, pp. R9.
DOI: 10.1186/gb-2012-13-2-r9.
Jain, A., Gali, H., and Kihara, D. (2018). Identification of
Moonlighting Proteins in Genomes Using Text Mining
Techniques. Proteomics, vol. 18, pp. 21-22.
doi:10.1002/pmic.201800083.
Jeffery, C. (2015). Why study moonlighting proteins? Front
Genet. 6:211. doi: 10.3389/fgene.2015.00211.
Jeffery, C. (2018). Protein moonlighting: what is it, and
why is it important? Philos Trans R Soc Lond B Biol
Sci., 373 (1738): 20160523. doi: 10.1098/rstb.2016.05
23.
Kawashima, S. and Kanehisa, M. (2000). AAindex: amino
acid index database. Nucleic Acids Res., vol. 28, pp.
374. doi: 10.1093/nar/28.1.374.
Khan, I., et al. (2016). Genome-scale prediction of
moonlighting proteins using diverse protein association
information. Bioinformatics, 32(15): 2281-2288.
doi:10.1093/ bioinformatics/btw166.
Mu, Z., et al. (2021). FEGS: a novel feature extraction
model for protein sequences and its applications. BMC
bioinformatics, vol. 22,1 297. doi:10.1186/s12859-021-
04223-3.
Piatigorsky, J., Wistow, G. (1989). Enzyme/crystallins:
gene sharing as an evolutionary strategy. Cell, 57(2):
197-9. doi: 10.1016/0092-8674(89)90956-2.
Ribeiro, D., et al. (2019) MoonDB 2.0: an updated database
of extreme multifunctional and moonlighting proteins.
Nucleic Acids Res., 47(D1): D398-D402. doi:
10.1093/nar/gky1039.
Shirafkan, F., et al. (2021). Moonlighting protein prediction
using physico-chemical and evolutional properties via
machine learning methods. BMC bioinformatics vol.
22,1 261. 24. doi:10.1186/s12859-021-04194-5.
Weaver, D. (1998). Telomeres: moonlighting by DNA
repair proteins. Curr. Biol. vol. 8, R492–R494. doi:
10.1016/S0960-9822(98)70315-X.
Predicting Moonlighting Proteins from Protein Sequence
275
