A Structure based Approach for Accurate Prediction of Protein

Interactions Networks

Hafeez Ur Rehman

, Usman Zafar

, Alfredo Benso

and Naveed Islam

1,3

Department of Computer Science, National University of Computer & Emerging Sciences, Hayatabad, Peshawar, Pakistan.

Department of Control & Computer Engineering, Politecnico di Torino, I-10129, Torino, Italy.

Department of Computer Science, Islamia College University, Peshawar, Pakistan.

Keywords:

Protein Interactions, Protein Structure, 3D Templates, Protein Interaction Network, Protein Binding Sites.

Abstract:

In the recent days, extraordinary revolution in genome sequencing technologies have produced an overwhelm-

ing amount of genes that code for proteins, resulting in deluge of proteomics data. Since proteins are involved

in almost every biological activity, therefore due to this rapid uncovering of biological “facts”, the ﬁeld of

System Biology now stands on the doorstep of considerable theoretical and practical advancements. Precise

understanding of proteins, specially their functional associations or interactions are inevitable to explicate how

complex biological processes occur at molecular level, as well as to understand how these processes are con-

trolled and modiﬁed in different disease states. In this paper, we present a novel protein structure based method

to precisely predict the interactions of two putative protein pairs. We also utilize the interspecies relationship

of proteins i.e., the sequence homology, which is crucial in cases of limited information from other sources

of biological data. We further enhance our model to account for protein binding sites by linking individual

residues in structural templates which bind to other residues. Finally, we evaluate our model by combining

different sources of information using Naive Bayes classiﬁcation. The proposed model provides substantial

improvements in terms of accuracy, precision, recall when compared with previous approaches. We report an

accuracy of 90% when tested for a protein interaction network of yeast proteome.

1 INTRODUCTION

Proteins are the most essential macromolecules that

are involved in almost every biological activity. Our

knowledge of new proteins is increasing with a rapid

pace as next generation sequencing technologies are

uncovering new genomes. The knowledge of proteins

alone, is not sufﬁcient since proteins rarely act in iso-

lation. The overall complexity of biological systems

at different levels primarily arise due to the combina-

torial interactions caused by the proteins in the cells.

One of the crucial step for understanding biological

cells as engineered systems is to map networks of

DNA-protein, RNA-protein and protein-protein inter-

actions (PPIs) of a species as completely and accu-

rately as possible. Precise knowledge of protein inter-

actions is also a precondition for fulﬁlling the promise

of preventive as well as personalized medicines that

which means more rational development of antibacte-

rial compounds, drugs, and vaccines.

The conventional wet lab experiments e.g., Yeast

two-Hybrid (Y2H) (Ito et al., 2001) screening,

Protein-fragment Complementation Assays (PCA)

(N. Pelletier et al., 1999), or co complex interaction

maps (that are attained by high-throughput Co-afﬁnity

Puriﬁcation followed by Mass Spectrometry (AP/MS)

to identify protein-protein (bait) interactions) (Rigaut

et al., 1999; A. Shoemaker and R. Panchenko, 2007a)

etc., are either slow, costly or prone to noise because

of the nature of these experiments. Moreover, the ex-

isting noise in protein interaction databases resulted

by these experiments, plus the deluge of protein data

produced by next generation sequencing technologies

motivates the need to make accurate computational

techniques that can precisely map the interactions of

proteins on genome wide scale.

Several computational techniques have been pro-

posed in the past that incorporate a wide variety

of data e.g., phylogenetic proﬁles, sequence homol-

ogy, and co-expression of genes etc., to accurately

infer genome-wide protein-protein interactions (A.

Shoemaker and R. Panchenko, 2007b; Salwinski and

Eisenberg, 2003). However, comparative studies ad-

vocate that the development of noise free protein in-

Rehman, H., Zafar, U., Benso, A. and Islam, N.

A Structure based Approach for Accurate Prediction of Protein Interactions Networks.

DOI: 10.5220/0005705002370244

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 3: BIOINFORMATICS, pages 237-244

ISBN: 978-989-758-170-0

 2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

237

Homolog Species

Saccharomyces

Cerevisiae Species

Hypothetical Protein

YKL033W-A

HDHD1 Protein

CG15441 Protein

Drosophila

Melanogaster

Homo Sapiens

Known Structures

Figure 1: A hypothetical protein connected to structurally

known proteins using protein homology.

teraction repertoires of different genomes, is still in

its early stages (Braun and et al., 2009; Deane et al.,

2002). The most prominent computational methods

that produce high conﬁdence interactions utilize pro-

tein’s structural information e.g., (C. Zhang et al.,

2010; Wass et al., 2011). But unfortunately, there is a

huge difference between the number of known protein

sequences and their relative known structures; even

for the well studied organism such as Saccharomyces

Cerevisiae, the known structural information is sparse

i.e. less than 10% proteins are with known structure

(Zhang et al., 2012). Moreover, the protein complex

information of known PPIs is even sparser.

Fortunately, homology models (see Figure 1) as

well as known protein complexes (across species) in

well-known databases e.g. PDB (Protein Data Bank)

(M. Berman et al., 2000), present the opportunity

to relate unknown structure sequences with known

structures using geometrical features of the individ-

ual templates. Approaches incorporating this type of

information have shown great success; in such cases

protein structure have multiple clues that associate

the geometric features of individual templates. How-

ever, these methods exhibit much less success on pro-

teins with inconsistent homolog templates (i.e. the

homolog templates whose geometrical features are

much variant; hence they result in effecting the over-

all accuracy of the prediction schemes).

In this paper we proposed and evaluated a novel

approach that combines heterogeneous structural in-

formation of proteins and determine their potency for

interaction in the form of a probability score. The fun-

damental conceptual innovation of our method is to

connect geometrical features of proteins with protein

binding sites and to enhance the algorithms power as

well as applicability for heterogeneous homolog tem-

plates. Our new approach relies on scores (features)

obtained by combining diverse sources of biological

information which includes: sequence similarity, pro-

tein homology, protein binding sites, and geometrical

features like, no of interacting residues, no of surface

residues etc. These scores are combined using Bayes

classiﬁer and an overall conﬁdence score is calculated

that determines the binding potential of two proteins

as interacting pairs.

The remaining part of the paper is organized as

follows: In Section 2, we ﬁrst give an overview of

the closely related approaches used for the predic-

tion of protein-protein interactions; with the expla-

nation of why structure based approaches stand out

from other techniques. We then introduce, in Section

3, a heterogeneous information based Bayesian net-

work model that combines different types of informa-

tion (i.e., sequence similarity, protein homology, pro-

tein binding sites, and geometrical features) to predict

protein-protein interactions. Section 4 demonstrates

the effectiveness of the proposed model when applied

to cross validate, a subset of interacting as well as non

interacting proteins in the yeast network. We lastly

discuss the results of our scheme and also compare

the performance of most recent related state-of-the-art

structure based schemes with our scheme. In Section

5 we present conclusion of our study with possible

future considerations.

2 RELATED WORK

Protein-protein interactions are key to most of the bio-

logical processes. These interactions orchestracted by

molecular mechanisms that have yet not been clearly

understood. Understanding protein-protein interac-

tions would also provide us crucial clues about in-

terracellular signalling pathways. Numerous experi-

ments have been devised by researchers in the labs

including yeast two-hybrid systems, mass spectrom-

etry, protein microarrays and others. Unfortunately,

experimental techniques have not been able to charac-

terize the proteins to a great extent. Thus, our knowl-

edge of protein functions as well as their interactions

is very limited. This low contribution by experimental

techniques and lesser knowledge about protein inter-

actions is being complemented by the advancement of

computational methods.

Since, protein sequence is the most basic as well

as most easily available type of information about

proteins; therefore, many methods devised in the be-

ginning focused on use of sequence information to see

the mutual evolution of proteins. One such method,

focused on evolutionary information related to struc-

ture and function was proposed by (Valencia and Pa-

zos, 2003). This method constructs and utilized evo-

lutionary relationship among proteins to infer PPI as

such proteins co-evolve. Another approach was a

multiple classiﬁer based system harnessing sequence

of proteins(F. Xia et al., 2010). They utilized two

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

238

classiﬁers rotation forest and autocorelation descrip-

tor. This group tested their system on Saccharomyces

cerevisiae and Helicobacter pylori data. Sequence

based approach for PPI prediction has been used by

another group but with slight variations. They pre-

dicted PPIs more precisely from sequence alignments

of proteins by using a Bayesian classiﬁer (Burger and

V. Nimwegen, 2008). A similar and more recent set of

techniques utilized only sequence information for PPI

prediction e.g., (Shen et al., 2006; You et al., 2015).

Mathematical probabilistic models were adopted

by some researcher for the prediction of protein-

protein interactions. In one such case Probabilitic

analysis predicted nearly 40,000 interactions in hu-

mans (R. Rhodes et al., 2005). This probabilistic

model combined interaction data, functional anno-

tation data, protein domain data and genome-wide

gene expression data. Probabilistic models have also

provided a motivation for researches to model more

protein-protein interactions. A work was done us-

ing Generative Probabilistic Models with bi clique

perspective to model the interaction network of Sac-

charomyces cerevisiae(Schweiger et al., 2011). This

method concluded that nave unmodiﬁed DD (dupli-

cationdivergence) model is much more effective than

Preferential Attachment model at capturing key as-

pects of PPI prediction. Another work employed the

use of distant conservation of patterns in protein se-

quences, also called motifs and their structural rela-

tionships in proteins (Espadaler et al., 2005).

Most recent approaches that integrate structural

and non-structural type of information into compu-

tational models and use machine learning algorithms

e.g., Bayesian classiﬁer or Support Vector Machines

etc., to infer interaction of putative proteins. One such

work is done in the recent past by (Zhang et al., 2012),

that utilizes structural as well as non-structural type of

features with a blend of Bayesian Classiﬁer for pre-

diction of PPI on a genome wide scale. The authors of

this study presented their results for Saccharomyces

cerevisiae, and reported that structural features out-

perform non structural features with great margin in

terms of statistical performance measures i.e., pre-

cision, recall, accuracy, false positive rate etc., The

major contribution in this work was the use of struc-

tural features and evaluation of their impact on the

prediction accuracy. Thus structural information of

proteins plays a key role in deciphering the underly-

ing mechanism of protein interactions. Therefore, in

our work we also mainly integrate, structural infor-

mation of proteins along with protein binding sites to

predict their associations.

3 METHODS

In our work, we employ the idea of integrating hetero-

geneous biological information associated with two

queried proteins and determine their strength for in-

teraction, by combining this information in the form

of scores using Bayesian statistics. The distinctive-

ness of our technique comes from the fact that po-

tential interaction information e.g., protein binding

sites (which are strongly associated to molecular in-

teraction), can be combined with geometrical features

present in the structural templates of two interacting

proteins to decide if they interact or not. This com-

bination also increases the power of our algorithm to

include structural templates that are varied in geom-

etry but contain sites that can bind to other proteins.

Our proposed approach relies on scores (features) ob-

tained by combining diverse sources of biological in-

formation which includes: sequence similarity, pro-

tein homology, protein binding sites, and geometrical

features like, no of interacting residues, no of surface

residues etc.

The prediction of protein interactions is more

challenging for proteins which are not well annotated

or whose molecular details are limited. To enhance

the predictive power of our automated PPI predic-

tion algorithm, we combine very powerful associative

sources of information namely: protein homolog &

sequence similarity, as a baseline to capture proteins

which are most similar. This is particularly impor-

tant as each type of data typically captures distinct

aspects of associative activity. The overall process of

our technique for PPI prediction is divided into seven

steps (as shown in Figure 02):

Step 1: Selection of Homolog Sequences

To predict the interactions for sparsely annotated pro-

teins, the ﬁrst useful type of information that can

associate them is the protein homology information.

Evolutionary relationships between species advocate

that the homolog (speciﬁcally orthologous) proteins

of different species, whose functions have been estab-

lished before speciation event and which share high

sequence similarity are likely to interact for similar

functional activities.

Two proteins are said to be homologous if they

share a common ancestor. To detect homology, se-

quence information is often used to deduce if proteins

are homologous or not. If two proteins share high se-

quence similarity i.e., above 25 % sequence similarity

(Benso et al., 2013; Mitrofanova et al., 2011; Benso

et al., 2012), they are very likely to be homologous

and have similar structures and in many cases part of

A Structure based Approach for Accurate Prediction of Protein Interactions Networks

239

BLAST

Calculate

Scores

on all 100

Template

Model

Pairs

Bayesian

Classifier

Yes/No

Decision

S2N1

NR Database

VAST+ Server

Multiprot

Program to

Detect

Interacting

Residue Pairs in

the Structure

PDBe Binding

sites server

For i, j Templates

i x j Template Pairs in

Total

Templates

Model by Superimposing

Templates with

Homologs

Finds Structural

Neighbor Templates

of each Homolog

Structure

. . . . . . . .

Query 1

Query 2

Homolog

PDB

Structures

Figure 2: The general shceme of heterogenous information integration for our PPI prediction algorithm.

the same molecular functional activity.

The input of our algorithm is a pair of proteins

(also called query proteins) say P

and P

(in our

implementation we used Uniprot IDs (The UniProt

Consortium, 2015)), whose interaction information

we want to ﬁnd or predict. In the ﬁrst step, since

these proteins can possibly be sparsely annotated so,

we need to associate them through homology to other

proteins. To capture ortholog based homolog similar-

ity we run a single iteration BLAST (Altschul et al.,

1990) search for each query protein P

against the pro-

tein NR database with an E-value cutoff of 0.0001.

We selected protein structures (namely, S

and S

;

also called Model structures) that are highly similar

to our query, with the additional constraint that match-

ing PDB structures should have at least 90% or higher

sequence identity. It is pertinent to note that this sim-

ilarity is only in sequences not in structure. The ob-

tained structures as a result of sequence similarity are

then queried to PDB (M. Berman et al., 2000), to ob-

tain their structural details (i.e. atomic coordinates,

residues information etc.).

Step 2: Finding Structural Neighbors

In the second step, structural representatives of each

model structure i.e. S

and S

, were taken directly

by querying each model structure to VAST+ (Vector

Alignment Search Tool Plus) Server (Madej et al.,

2013). VAST+ is a tool designed by NCBI (Na-

tional Center for Biotechnology Information) and uti-

lizes Molecular Modeling Database (MMDB), for 3-

dimensional structures, with the need of ﬁnding those

structures that have similar macromolecular com-

plexes. The macromolecular similarities are evalu-

ated using purely molecule’s geometric criteria, with-

out considering sequence similarity, thus it is able to

identify even distant homologs structures. We queried

VAST+ with default parameters and with a thresh-

old of ten templates i.e. we select top ten neighbor-

ing templates of each model structure. The structural

neighbors are named as SA

for model template S

and SB

for model template S

, (where, i = 1,2.....10)

Step 3: Formation of Templates Pairs

At this stage, we have 10 structural neighbors for each

query protein P

. To check the overall binding poten-

tial for interaction, of individual template pairs, we

construct pairs of each structural neighbor SA

with

(where i = 1,2.....10) i.e., SA

pairs with SB

.... and so on up to SB

, likewise we repeat pair-

ing for SA

, SA

.... up to SA

. This step results in a

total of 100 template pairs.

Step 4: Identiﬁcation of Interacting

Residues and Binding Sites

As a ﬁrst step, to evaluate the propensity for inter-

action, of individual template pairs, we ﬁrst identify

the # of interacting residues in the template pairs. For

this purpose, we use a tool called Multiprot in a proto-

col known as PRISM (PRotein Interactions by Struc-

tural Matching) (Tuncbag et al., 2011; Shatsky et al.,

2004) . The Multiprot rationale is based on the fact

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

240

that globally different protein structures can interact

via chains of architecturally similar residues called

motifs. Thus Multiprot predicts binding residues by

utilizing structural similarity as well as evolutionary

conservation of putative binding residue also called

hot spots. For each template pair Multiprot calculates

the # of interacting residues.

To further strengthen the PPI prediction of our

technique we also utilize the PDBeMotif (Golovin

and Henrick, 2008). PDBeMotif is an incredibly fast

and powerful search tool that facilitates the explo-

ration of binding sites of single proteins or classes of

proteins e.g., Pepsin, and locates the conserved struc-

tural features of individual residues both within the

same specie as well as in different species. We em-

ploy PDBeMotif to locate residues that are binding

sites in our template pairs.

Step 5: Modeling Structural Templates

using Homolog Pairs

In this step, we build an interaction model Mod

superposing the template pairs SA

and SB

over the

model template S

and S

. Overall 100 models are

built for (10x10) template pairs. Each model Mod

used to calculate four structure based scores.

Step 6: Calculating Interaction Scores

from Interaction Models

From the 100 interaction models we prepared in the

previous step, we evaluate and combine associated in-

formation to calculate four scores for each interaction

model Mod

. The scores are based on the criterion

that make use of interacting residues, binding sites

as well as sequence information. We name our ﬁrst

score as ξ

(1)

Mod

i j

, where Mod

denotes the interaction

model for which this score is calculated. ξ

(1)

Mod

i j

calculated by taking into account the number of in-

teracting residues in the template (calculated using

Multiprot) that are preserved in the homolog mod-

els S

and S

, i.e. both template and model share

those residue pairs. Templates have different varia-

tions in their amino acid sequence, this score captures

the strength of interaction model in terms of # of in-

teracting residues preserved, when compared with ho-

molog template pair.

The second score of our model is called ξ

(2)

Mod

i j

and is estimated by taking fraction of total interacting

residues preserved i.e., ξ

(1)

Mod

i j

, divided by the average

of total number of residues in both homolog templates

i.e., S

and S

, as shown in equation 1.

(2)

Mod

i j

(1)

Mod

i j

Average(S

)

(1)

The third score ξ

(3)

Mod

i j

is the same as the ﬁrst score,

with the additional check that the interacting residues,

both in template and model are also shared by the

binding sites retrieved using PDBeMotif service, and

is calculated as shown in the equation 2.

(3)

Mod

i j

(1)

Mod

i j

∩ Binding

−

Sites(Mod

i j

)

(2)

Lastly, the ﬁnal score ξ

(4)

Mod

i j

, of our technique is

calculated by taking shared binding sites in the super-

imposed template and model pairs as shown in equa-

tion 3. ξ

(4)

Mod

i j

is the number of binding sites in the

template that align to the number of binding sites in

the model.

(4)

Mod

i j



Binding

−

Sites(S1, S2) ∩ Binding

−

Sites(Mod

i j

)



(3)

Step 7: PPI Prediction using Bayesian

Networks

Once all scores are calculated for hundred interaction

models, we then combine their effect into one score

by taking the mean and standard deviation of individ-

ual scores as shown in equation 4 and 5.

(k)





∑

i=1

∑

j=1

(k)

Mod

i j

100





.. .For, k = {1, 2,3,4}

(4)

(l)







∑

i=1

∑

j=1



(k)

Mod

i j

− ϕ

(k)



100







.. .For, k = {1, 2,3,4} and l = {5,6, 7,8} (5)

The Standard deviation of scores captures the fact

that, whether the templates that our method ﬁnds are

different from each other or not; because when dif-

ferences among homologs are spread out the standard

deviation will be high.

Lastly we use Bayesian classiﬁcation to com-

bine the mean values as well as the standard de-

viations of our scores captured in eight variables

(k)

, where k={1,2,...,8}. Let Pi and Pj be the

query proteins whose interaction we want to predict

A Structure based Approach for Accurate Prediction of Protein Interactions Networks

241

and ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

be the random vari-

able that capture different aspects of structural as-

sociation. The conditional probability that Pi and

Pj interact given the distribution of random variables

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

is given by:

P(C

i j

= 1/ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

) =

P(ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

i j

= 1).P(C

i j

= 1)

P(ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

)



k=1

P(ϕ

(k)

i j

= 1).P(C

i j

= 1)





k=1

P(ϕ

(k)

i j

= 1).P(C

i j

= 1)





k=1

P(ϕ

(k)

i j

= 0).P(C

i j

= 0)



(6)

Where P(C

i j

= 1) is the prior probability that

Pi and Pj interact, P(ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

)

is the probability that Pi and Pj has

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

features and

P(ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

i j

= 1) is the proba-

bility that Pi and Pj has ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

features given that Pi and Pj interact.

All feature values ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

are

normalized and we used binning of feature values

so that values of features lie in known ranges. As

many machine learning algorithms specially Bayes

classiﬁcation produce better results when continuous

attributes are made discrete. Finally, we calculate

the value of P(C

i j

= 1/ϕ

(1)

,ϕ

(2)

,ϕ

(2)

,.........,ϕ

(8)

) for

each protein pair Pi and Pj.

4 EXPERIMENTAL SETUP AND

RESULTS

The integration technique described in the Methods

section is evaluated on the task of predicting protein-

protein interactions for an interaction network of Sac-

charomyces cerevisiae species proteins. We tested

our algorithm on a data set of Yeast species proteins

obtained from IntAct database [Results-01]. The al-

gorithm fuses probabilities derived from diverse data

sources including sequence similarity, protein homol-

ogy, protein binding sites used in combination with

other geometrical features. A well known power-

ful classiﬁcation scheme i.e. Bayes classiﬁcation,

was used to combine mutually independent features

(scores).

In this work, we present results of our scheme

for a portion of Saccharomyces cerevisiae species

interaction network. We chose the interaction net-

work of HSP75 YEAST protein (Uniprot ID: P11484)

for our experiment and tried to reproduce its inter-

action network using our proposed algorithm. The

HSP75 YEAST protein was chosen ﬁrstly because it

is involved in heterogeneous molecular activities and

secondly, because the interaction networks of this

protein contains many experimentally validated in-

teractions. Thus, to better evaluate the prediction

performance of our algorithm we chose this net-

work. The HSP75 YEAST is a fully reviewed protein

in UniProtKB/Swiss-Prot database (which is a high

quality manually curated, as well as non-redundant

protein sequence database).

The HSP75 YEAST protein’s interaction network

in IntAct database contain 4,449 interactions as of

August, 2015. The protein interaction network

databases contain false positive interactions that are a

bottleneck to predict the overall performance of an al-

gorithm as well as to judge the statistical signiﬁcance

of experiments conducted. In order to deal with this

limitation, we ﬁltered interaction network to include

interactions that are of high conﬁdence with the crite-

ria that each interaction in the network must be sup-

ported by at least two experimental methods. After

ﬁltering our network reduced to 1770 interactions.

We call these interactions as high conﬁdence in-

teractions because each interaction is supported and

validated by at least two experimental methods. The

interaction network contains proteins from the same

(Saccharomyces cerevisiae) as well as other species

namely: Arabidopsis thaliana, Rattus norvegicus,

Arabidopsis thaliana, and Dictyostelium discoideum.

4.1 Performance Evaluation

For evaluating prediction performance we use cross

validation approach to estimate the prediction potency

of our proposed scheme i.e., for each protein pair Pi

and Pj in the interaction data set, we assumed the

interaction of Pi and Pj were unknown and then at-

tempted to predict the interaction by means of our

algorithm. Lastly, we compare the predicted inter-

actions with the true interaction set. For assessment

of our methodology, we computed performance mea-

sures, such as: precision, recall, accuracy and F1

which are estimated using the following formulas:

accuracy =

T P + T N

T P + FP + T N + FN

precision =

T P

T P + FP

recall =

T P

T P + FN

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

242

and

F1 =

2 ∗ precision ∗ recall

precision + recall

4.2 Cross Validation Analysis of

Prediction Accuracy

For the interaction network described earlier we ﬁrst

attempted to predict protein-protein interactions by

10-fold cross validation. For each protein pair Pi and

Pj the probability that protein Pi interacts with Pj is

calculated using equation 6. Predicted protein interac-

tions having a probability estimate of greater than 0.5

were considered as positive interactions otherwise we

conclude that proteins don’t interact. By applying our

algorithm on high conﬁdence interaction network re-

trieved from IntAct, we obtained an overall accuracy

of 90%, recall of 95.2%, precision of 94.11 and an F1

score of 94.49%.

4.3 Comparison with other Approaches

In this section, we broadly compare our method to the

most widely used group of techniques, such as Pre-

PPI algorithm proposed by Q. C. Zhang et al. (Zhang

et al., 2012), which combines structural as well as

non structural type of information to predict protein-

protein interactions. In such methodologies, inter-

actions among proteins are predicted by combining

structural clues with non structural clues using some

machine learning algorithm such as, Support Vector

Machines (SVM), Bayesian framework etc., which

consequently assign a probability score to a protein

pair of interest as positively or negatively interact-

ing. Fundamentals of Bayesian techniques are at the

heart of the overwhelming majority of methods cur-

rently used to combine heterogeneous sources infor-

mation for PPI prediction. Since this scheme (Zhang

et al., 2012), uses Bayesian technique as well as uti-

lizes structural information to predict PPI therefore,

we compare our algorithm against this computational

technique.

To obtain the most correct comparative results,

we use the same species proteins i.e., Saccharomyces

cerevisiae and compare results in a 10 fold cross-

validation setting. The results in ﬁgure 03 clearly sig-

nify that our method performs better than the Q. C.

Zhang’s Pre-PPI method (Zhang et al., 2012) across

all measures reported i.e., precision, recall, accuracy

and F1 scores. We observed that for almost the same

accuracy values, Pre-PPI method produced higher

number false positive as well as false negatives pre-

dictions, which resulted in lower values of precision

Acurracy

Recall

Precision

Our Method

95.2

94.11

94.49

Pre-PPI Method

100

Percentage of values corresponding to each measure

Comparison with Pre-PPI

Figure 3: Comparison of Acurracy, Recall, Precision, and

F1 measure of proposed scheme with Pre-PPI scheme.

and recall, respectively. The improved performance

of our algorithm can be attributed to the most impor-

tant functional clue called protein binding sites, which

was further improved by combining with other struc-

tural information to precisely model the interaction

activity.

5 CONCLUSIONS

In this work, we presented a novel approach that uses

heterogeneous biological information associated with

two queried proteins and determine their strength for

interaction, by combining this information in the form

of scores using Bayesian statistics. The distinctive-

ness of our technique comes from the fact that po-

tential interaction information i.e., protein binding

sites, can be combined with other geometrical fea-

tures present in the structural templates of two inter-

acting proteins to decide if they interact or not. This

combination also increases the power of our algo-

rithm to include structural templates that are varied

in geometry but contain sites that can bind to other

proteins. The proposed model provides substantial

improvements in terms of accuracy, precision, recall

when compared with previous approaches. The pro-

posed scheme may additionally be used in combina-

tion with non structural features to enhance the pre-

diction conﬁdence.

ACKNOWLEDGEMENTS

We would like to show our gratitude to Dr. Omar

Usman, Assistant Professor at National University of

Computer & Emerging Sciences for his worthy com-

ments that greatly improved the manuscript.

A Structure based Approach for Accurate Prediction of Protein Interactions Networks

243

REFERENCES

A. Shoemaker, B. and R. Panchenko, A. (2007a). De-

ciphering protein-protein interactions. part i. exper-

imental techniques and databases. PLOS Comput.

Biol., 3(3):e42.

A. Shoemaker, B. and R. Panchenko, A. (2007b). Deci-

phering protein-protein interactions. part ii. computa-

tional methods to predict protein and domain interac-

tion partners. PLOS Comput. Biol., 3(3):e43.

Altschul, S., Gish, W., Miller, Myers, E., and J. Lipman, D.

(1990). Basic local alignment search tool. Molecular

Biology, 215:403–410.

Benso, A., Di Carlo, S., Ur Rehman, H., Politano, G.,

Savino, A., and Suravajhala, P. (2012). Using gnome

wide data for protein function prediction by exploit-

ing gene ontology relationships. pages 497–502. IEEE

International Conference on Automation Quality and

Testing Robotics (AQTR)., IEEE.

Benso, A., Di Carlo, S., Ur Rehman, H., Politano, G.,

Savino, A., and Suravajhala, P. (2013). A combined

approach for genome wide protein function annota-

tion/prediction. PROTEOME SCIENCE, 11(S1):1–

12. ISSN: 1477-5956.

Braun, P. and et al. (2009). An experimentally derived con-

ﬁdence score for binary protein-protein interactions.

Nature Methods, 6:91 to 97.

Burger, L. and V. Nimwegen, E. (2008). Accurate pre-

diction of protein protein interactions from sequence

alignments using a bayesian method. Mol Syst Biol,

4:165.

C. Zhang, Q., Petrey, D., Norel, R., and Honig, B. (2010).

Protein interface conservation across structure space.

Proc. Natl Acad. Sci. USA, 107:10896–10901.

Deane, C. M., Salwinski, L., Xenarios, I., and Eisenberg,

D. (2002). Protein interactions: two methods for as-

sessment of the reliability of high throughput observa-

tions. . Mol. Cell. Proteomics, 1:349 to 356.

Espadaler, J., Romero, O., M. Jackson, R., and et al. (2005).

Prediction of protein-protein interactions using distant

conservation of sequence patterns and structure rela-

tionships. Oxford Journals, Volume 21, Issue 16:3360

–3368.

F. Xia, J., Han, K., and S. Huang, D. (2010). Sequence-

based prediction of protein-protein interactions by

means of rotation forest and autocorrelation descrip-

tor. Protein Pept Lett, 17(1):137–45.

Golovin, A. and Henrick, K. (2008). Msdmotif: exploring

protein sites and motifs. BMC Bioinformatics, 9:1–11.

Springer-Verlag Berlin Heidelberg.

Ito, T., Chiba, T., Ozawa, R., and et al. (2001). A com-

prehensive analysis of protein protein interactions in

saccharomyces cerevisiae. Proc Natl Acad Sci USA,

98:4569–74.

M. Berman, H., Westbrook, J., Feng, Z., Gilliland, G.,

N. Bhat, T., Weissig, H., N. Shindyalov, I., and E.

Bourne, P. (2000). The protein data bank. Nucleic

Acids Research, 28:235–242.

Madej, T., J. Lanczycki, C., Zhang, D., A. Thiessen,

P., C. Geer, R., M. Bauer, A., and H. Bryant, S.

(2013). Mmdb and vast+: tracking structural simi-

larities between macromolecular complexes. Nucleic

Acids Res., 42:(D1): D297–D303. [PubMed PMID:

24319143].

Mitrofanova, A., Pavlovic, V., and Mishra, B. (2011). Pre-

diction of protein functions with gene ontology and in-

terspecies protein homology data. IEEE/ACM Trans-

actions on Computational Biology and Bioinformat-

ics, 8 no. 3:775–784.

N. Pelletier, J., Arndt, K., Pluckthun, A., and et al. (1999).

An in vivo library versus library selection of opti-

mized protein protein interactions. Nat Biotechnol,

17:683–90.

R. Rhodes, D., A. Tomlins, S., and Varambally, S. (2005).

Probabilistic model of the human protein-protein in-

teraction network. Nature Biotechnology, 23:951 –

959.

Rigaut, G., Shevchenko, A., Rutz, B., and et al. (1999). A

generic protein puriﬁcation method for protein com-

plex characterization and proteome exploration. Nat

Biotechnol, 17:1030–32.

Salwinski, L. and Eisenberg, D. (2003). Computational

methods of analysis of protein protein interactions.

Curr. Opin. Struct. Biol., 13:377 to 382.

Schweiger, R., Linial, M., and Linial, N. (2011). Gener-

ative probabilistic models for protein-protein interac-

tion network the biclique perspective. Oxford Jour-

nals, Volume 27.

Shatsky, M., Nussinov, R., and J. Wolfson, H. (2004). A

method for simultaneous alignment of multiple pro-

tein structures. PROTEINS: Structure, Function, and

Bioinformatics, 56:143–156.

Shen, J., Zhang, J., Luo, X., Zhu, W., Yu, K., and et al.

(2006). Predicting protein-protein interactions based

only on sequences information. Proceedings of the

National Academy of Sciences, vol. 104:4337–4341.

The UniProt Consortium (2015). Uniprot: a hub for protein

information. Nucleic Acids Res. 43: D204-D212.

Tuncbag, N., Gursoy, A., Nussinov, R., and Keskin, O.

(2011). Predicting protein-protein interactions on a

proteome scale by matching evolutionary and struc-

tural similarities at interfaces using prism. Nature Pro-

tocols, 06 NO.09:1341–1354.

Valencia, A. and Pazos, F. (2003). Prediction of protein-

protein interactions from evolutionary information.

Methods Biochem Anal, 44:411–26.

Wass, M., Fuentes, G., Pons, C., Pazos, F., and Valencia,

A. (2011). Towards the prediction of protein interac-

tion partners using physical docking. Mol. Syst. Biol.,

7:469.

You, Z. H., Chan, K. C. C., and Hu, P. (2015). Predict-

ing protein-protein interactions from primary protein

sequences using a novel multi-scale local feature rep-

resentation scheme and the random forest. PLoS ONE,

10(5).

Zhang, Q. C., Petrey, D., and et al. (2012). Structure based

prediction of protein-protein interactions on a genome

wide scale. Nature, 490(7421):556 to 60.

BIOINFORMATICS 2016 - 7th International Conference on Bioinformatics Models, Methods and Algorithms

244