On the Impact of Granularity in Extracting Knowledge from

Bioinformatics Data

Sean West and Hesham Ali

College of Information Science and Technology, University of Nebraska at Omaha, Omaha, U.S.A.

Keywords: Data Integration, Knowledge Extraction, Gene Expression Data, Protein-protein Interaction, Co-regulation,

Correlation Networks, and Clusters.

Abstract: With the rapidly increasing amount of various types of biological data currently available to researchers, the

focus of the biomedical research community has been shifting from pure data generation towards the

development of new methodologies for data analytics. Although many researchers continue to focus on

approaches developed for analyzing single types of biological data, recent attempts have been made to utilize

the availability of heterogeneous data sets that contain various types of data and try to establish tools for data

integration and analysis in many bioinformatics applications. Such attempts are expected to increase

significantly in this coming decade. While this can be viewed as a positive step towards advancing big data

analytics in bioinformatics, it is critical that these integration methodologies are meticulously studied to

ensure high quality of the knowledge extracted from the integrated data. In this work, we employ data

integration methods to analyze biological data obtained from protein interaction networks and gene expression

data. We conduct a study to show that potential problems can arise from integrating or fusing data obtained

at different granularity levels and highlight the importance of developing advanced data fusing techniques to

integrate various types of biological data for analytical purposes. Further, we explore the impact of granularity

from a more formulized approach and the granularity levels significantly impact the quality of knowledge

extracted from the integrated data.

1 INTRODUCTION

The bioinformatics perspective of data integration is

the uncovering of biological data and the extraction

of useful biological information (Rhee et al., 2008).

With the subsequent push towards data aggregation

and integration (Chatr-aryamontri et al., 2013;

Salwinski et al., 2004), comes a series of challenges,

highlighted by rapidly changing bioinformatics data

standards (Prasad et al., 2009; Kerrien et al. 2011).

Many of these bioinformatics data standards are

suitable for aggregation by targeting data reporting

and storage, such as the Minimum Information about

a high-throughput SEQuencing Experiment for

microarrays (Ceol et al. 2009), standards of data-use

are influenced by research outcomes and must be

more flexible to handle the swift evolution of

community ordained workflows. Particularly,

standards must handle the sensitivities of data sources

within these evolving workflows.

Data fusion is a special case of data integration

where two or more pieces of data are combined to

create a new parameter with its own novel meaning.

Although the term data fusion is relatively new to

bioinformatics, long associated with a military

connotation, its utilization is becoming increasingly

popular, with 21 PubMed publications in 2005 using

the term “data fusion”, and 95 publications in 2015.

Data fusion is a multi-step process, cascading

from the primary step of data source selection

(Taneera et al., 2012; Hanisch et al. 2002). The high

complexity of bioinformatics data sources creates a

special challenge (Bossi and Lehner, 2009). Not only

does this complexity enhance the central

characteristics of big data, such as variety and

veracity, but it accentuates the problem of

granularity.

Granularity refers to the shifts in scale where

membership is defined through mereology (Bittner

and Smith, 2003) or indiscernibility (Hobbs, 1995).

These two granularity dimensions were originally

specified as abstraction, shifts in specificity, and

aggregation, shifts in part-whole relations (McCalla

et al. 1992). Later the aggregation dimension was

adapted to into granularity parthood, molecules in a

cell, and determinate parthood, functioning members

West, S. and Ali, H.

On the Impact of Granularity in Extracting Knowledge from Bioinformatics Data.

DOI: 10.5220/0005778700920103

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 3: BIOINFORMATICS, pages 92-103

ISBN: 978-989-758-170-0

of the cell (Bittner et al., 2004). These scales have not

seen a lot of change in recent publications, and the

term granularity usually attributed to specificity.

However, increased differentiation of granularity

scales have been specified for ontology purposes

(Rector et al., 2006; Vogt et al., 2012). In the

biomedical domain, data sources contain members

from a population of available data-producing

sensors, we refer to the determinate parthood scale in

this study. Further, we use the terms abstraction and

aggregation in reference to their associated

dimensions of granularity, instead of their traditional

data processing definitions.

When examining biomedical data sources within

the abstraction dimension, two overarching

categories arise, isolated and integrated data sources.

Isolated data sources are typically specific,

representing results from single experiments.

Integrated data sources, if single-modal, may exist at

a similar abstraction level as isolated data sources.

However, if multi-modal, due to multiple

technologies being implemented, the data will exist at

a higher abstraction and consequently lower

granularity level.

We use microarray data as an example of low

abstraction, high granularity data, since each series

usually represents just a few experimental conditions

across a limited number of tissues. The variability of

cellular function within these tissues necessitates that

microarray data is not the epitome of high granularity

data, rather it exists at a granularity level where

differences between cellular conditions can be

extracted. To combat high false-positive and false-

negative rates, microarray is often enriched through

low-granular domain knowledge. A key component

to microarray data analysis is to differentiate between

cellular conditions. One data fusion methodology is

to put these differences in the context of domain

knowledge as a component of network creation

(Agarwal et al., 2008), through enrichment (Xu et al.,

2011), or examining expression differences within

the protein-protein interaction (PPI) network

(Medintz et al., 2007).

Protein-protein interaction databases may contain

high abstraction, low granularity data. Some recent

PPI databases are cell-specific or even molecule

specific (Veres et al., 2014; Liu et al., 2011).

Additionally, many integrated PPI databases, such as

the Search Tool for the Retrieval of INteracting

Genes/Protein (STRING), compile a list of potential

relationships, not taking unique cellular conditions

into account. STRING scores the interaction between

proteins across a set of data sources in a union-like

fashion (Bindea et al. 2009). Here, we use PPI data

sources that are non-integrated and not condition

specific, in order to bias the data towards a low

granularity. These non-integrated PPI databases use

manual curation methods to extract PPI information

from scientific literature. So, even non-integrated PPI

data sources are examples of multi-modal systems.

Yet, since the manual curation methodologies

employed to create PPI databases may innately

increase the granularity of the data, the diversity in

the curation methods may lead to the lower-

abstraction levels. In this work, we examine the

structural and biological attributes of several popular

PPI databases in order to characterize their unique

contributions towards data integration. We further

examine their pathway enrichment of each database

to determine any specificity or unique bias towards

similar groupings of biological functionality, which

would indicate increased levels of granularity.

Although the differences may be explicit between

cellular conditions from the expression data and since

PPI data comes from high abstraction data sources,

integrating microarray data with PPI data that is not

tissue or cellular condition specific does not model

the true protein-protein interaction network within the

experimental cellular condition. Therefore, the

consequences of alternate expression and PPI

network structure changes may not depict true

biological reality. If the variability in granularity

levels between PPI databases and microarray data

biases the data away from high-granularity,

potentially questionable biological information will

be extracted after the data fusion implementation. To

test this critical point, we can fuse the PPI and the

microarray data and compare the information

extraction between the original experimental

microarray data and the fused datasets. In this study,

we test to see the effect of fusing low abstraction,

microarray data with high abstraction, PPI data on

extraction of Type II Diabetes specific pathways.

Granularity along the aggregation dimension

requires a more formulized definition, which is

specified within the methods section. This

formulization includes three suggestions for

aggregation definition. First, using a rough set theory

definition of granularity, data fusion of specific data

sources lies on an abstraction granularity level

dependent on a set of attributes governing the

differentiation between data sources. Alternate

abstraction levels along scales defined by a set of

relevant attributes, impact the results of biological

information extraction by biasing the fusion towards

those chosen attributes. Second, when separating

these data sources according to attributes we create a

set of fusion networks for each attribute set used. The

On the Impact of Granularity in Extracting Knowledge from Bioinformatics Data

segmentation of the original data sources can also be

used to define level of aggregation. Finally, the

number of original data sources can be used to define

aggregation within each fused network separately.

We use these three definitions of aggregation to

test the relationship between information extraction

and aggregation. So, in summary, we have three

hypothesis:

H1: The different curation methods of non-

integrated PPI databases do not offer unique

bias towards specific biological functionality.

H2: Fusion of low-abstraction and high-

abstraction data will decrease experimental-

specific information extraction.

H3: There exists a definition of aggregation such

that a relationship between granularity and

information extraction can be seen.

Section 2 describes the methods for this study,

including the formulizations for the definition of

granularity in the aggregation dimension. Section 3

depicts the results. Section 4 discusses the outcome of

the study and its impact on the hypotheses. The paper

concludes with section 5.

2 METHODS

Throughout the study, protein-protein interaction data

and microarray data are modeled as networks, where

the nodes represent the biological elements and the

edges connect elements that are related by interaction

or high correlation. In the first part, we attempt to find

unique biological themes or functionalities associated

with the PPI databases queried in order to answer

hypothesis H1. We use structural similarity between

the PPI networks and the quality of the clusters

obtained from the networks using standard pathway,

disease, and ontology enrichments. In this manner,

we identify the biological functionality associated

with each protein-protein interaction network and

enriched clusters are mapped to human pathway

hierarchies to search for significant patterns.

In the second part, we address hypothesis H2. To

test the hypothesized relationship between

abstraction and data extraction, we use a case study

with a Type II Diabetes microarray series. We create

the integrated network using PPI and microarray data.

We then enrich obtained network clusters to identify

network-specific biological functions. We choose a

list of 24 diabetes associated pathways or diseases

curated from Reactome, the Online Mendelian

Inheritance in Man (OMIM), and the Kyoto

Encyclopedia of Genes and Genomes (KEGG). We

assess the enrichments of these pathways across the

original and fused networks, and discuss the potential

information loss that may occur due to the lack of

consistent granularity levels.

To validate the relationship between aggregation

granularity and knowledge extraction, hypothesis H3,

we formulize granularity using three different

approaches and add additional sources to expand on

the number of discrete granularity levels that can be

Figure 1: Data sources of this study. Stage one corresponds to hypothesis 1 and uses only the PPI databases. Stage 2 uses a

selection from the PPI databases and conducts an enrichment comparison. Stage 3 uses all the data sources in an information

extraction test to understand its interplay with aggregation.

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

formed. Extraction scores based on enrichment

results are assigned to each network and correlation

is measured.

2.1 Protein-protein Interaction

Databases

The following protein-protein interaction databases

were selected for this study since they reflect

variability associated with experiments used to obtain

them, they do not have a high-degree of integration

among each other, and they were initially, seemingly

sources of low granularity. The databases used were

Database of Interacting Proteins (DIP) (Thorne and

Stumpf, 2007), BioGRID (Zhang and Horvath, 2005),

Human Protein Reference Database (HPRD)

(Obayashi and Kinoshita, 2009), IntAct (Ingram et

al., 2006), and Molecular INTeraction database

(MINT) (Kashani et al., 2009).

DIP focuses on extracting experimental

knowledge from publications and stores binary

interactions between proteins, clarifying source and

evidence. BioGRID is a database of protein and

genetic interactions that are extracted from manually

annotated publications, by a team of PhD curators.

Text mining is used to rank relevant publications

where interactions are manually extracted and added

to BioGRID. HPRD uses laboratory submitted data

through a tool called BioBuilder which helps

researchers interact with the database and submit

experimental information. In this way, HPRD has

protein-protein interactions that are post-translational

modification, disease, and tissue specific. It also has

an overarching binary PPI source. IntAct takes an

open-source approach, with all data and repository

code available to the public. The stored interactions

are publicly curated from literature but also have a

design to allow for direct researcher annotation. Rules

on curation are specified on the EBI website and

interactions are reviewed by a second curator. MINT

is highly similar to IntAct, using the same

infrastructure and curation rules. The difference is the

set of MINT curators.

For the third (aggregation) component of the

study, we expand on those PPI databases with

curation conducive to tissue or disease specificity.

We add a diabetes sub-network for IntAct. HPRD has

many tissue specific curations, but we use only the

HPRD subnetworks with attribute overlap for the

microarray series used in the third part. So, we

include skeletal muscle, β-cell, pancreas, and blood

HPRD PPI networks.

2.2 Network Creation

Although the protein-protein interaction databases

contained evidence codes which may affect edge

weights through confidence variance, the granting of

specific edge weights was not implemented. This

alleviated the necessity of consolidating the PPI edge

weights with the microarray edge weights. Instead,

edges exist where evidence supports an edge. Further,

all types of experiments, including high-throughput

evidences, were included if they were present in the

original PPI data source. This may introduce a

technology bias beyond what is incorporated into the

research bias. However, correction of a technology

bias may introduce unknown sensitivities. So,

networks created were binary and non-directional.

Protein-protein interaction networks were derived

from the overarching sets of database information,

such that tissue specific information was included

without its specificity. Only complete proteins which

correspond to at least one Ensembl gene Id were

utilized. We attempt to highlight the issues of

removing granularity from domain knowledge

sources. Yet for validating the concept of the

interaction between aggregation and information

extraction, we use PPI networks with higher levels of

granularity as mentioned above.

Microarray data was initially downloaded from

the Gene Expression Omnibus series, GSE 38642

(Halevy et al., 2006). This series was chosen since it

is human, has a large set of biological replicates,

demonstrates a disease with a long list of well-

characterized pathways, Type II Diabetes, and

obtains expression through a relevant and specific

tissue, pancreatic islets. Additionally for the third part

of the study, we included series GSE 30803, a

treatment based study on healthy β-cells, GSE 67297,

a study on cold acclimation effects of diabetic adipose

tissue, GSE 55100, a blood tissue study of diabetes,

and GSE 59363, which uses skeletal muscle tissue in

healthy and diabetic samples with exercise stages.

These additional microarray series were chosen as

they have at least a moderate number of biological

replicates, and overlapping values across “tissue”,

“disease state”, “treatment”, and “technology”

attributes.

The raw expression files were downloaded, and

robust multi-array (RMA) normalized. Pearson

correlation was implemented to find expression

relationships. The microarray networks then took two

different paths, those filtered through false-discovery

rate p-value correction and those with hard thresholds

at 0.8 power and a 0.05 p-value. Base mapping of

probes to Ensembl gene Ids was completed through

On the Impact of Granularity in Extracting Knowledge from Bioinformatics Data

the Biomart API. Ensembl gene Ids which correspond

to multiple probes were assigned edge weights which

matched the highest correlation scoring probe for

each individual interaction. The strong influence of

some protein domains (e.g. probes which correspond

to multiple transcripts) reduce the accuracy of the

correlations which use the probe’s expression values.

The conjugated expression values are representative

of multiple transcripts. Depending on the abundance

distribution of these transcripts either correlations

may be assigned to the wrong protein, or more likely,

the correlations will favor random correlation values,

which are more likely to be insignificant and

negligible. These multi-transcript probes are

considered negligible in this study as they make up

only a small percentage of the total probes.

With the lack of PPI exact interaction strength

values, the integrated networks created were the

union of the PPI matrices and the microarray

matrices. Union is a surprisingly common kernel

function when integrating and fusing biological data

sources. We use it here as an example of an

integration-based data fusion approach.

2.3 Identification of Unique

Contributions from Protein-protein

Interaction Data and Type II

Diabetes Case Study

PPI networks were clustered with the Speed and

Performance In Clustering (SPICi) algorithm, a fast

and biologically driven clustering approach (Jiang

and Singh, 2010). The standard parameters produced

ideally sized clusters for enrichment. An in-house

tool for enrichment which downloads source groups

and group information for Reactome, OMIM, and

KEGG datasets. It uses the multivariate

hypergeometric function to find overly expressed

source groups within network clusters. Then, it uses

the Benjamini-Hochberg-Yekutieli false discovery

rate p-value correction to address multiple hypothesis

testing and dealing with the lack of independence for

enrichment terms on a single cluster.

Unique contributions were determined by finding

those enrichments for a PPI source that were not

identified in any other PPI source. For visualization,

unique Reactome enrichments were mapped to the

Reactome pathway hierarchy and grouped by

pathway similarity. Further, structural differences

between PPI networks were uncovered at the node,

edge, and cluster levels.

The microarray networks were fused with the PPI

networks in a union fashion so that there were control

microarray, diabetes microarray, PPI, control fused,

and diabetes fused networks. These networks were

filtered as to only include only those biological

elements present in the microarray sets. Then they

were clustered and enriched using SPICi and the in-

house enrichment tool. Diabetes pathways were

manually determined for Reactome, GO, OMIM, and

KEGG. Enrichments of these pathways were

examined across the networks to identify biological

differences between control and diabetes networks.

2.4 Validation of Relationship between

Aggregation and Information

Extraction

A more formulized definition of granularity is

required to characterize the relationship between

granularity and information extraction. So far, we use

the dimension of abstraction. This allows only for

direct comparisons between objects or networks

along the same scale. An extended discrete

comparison scale is needed along the dimension of

aggregation.

Shortly after the initial introduction of rough sets

into uncertainty theory (Pawlak, 1982), Hobbs began

to distinguish granularity as a significantly

contributing factor towards uncertainty (Hobbs,

1985). This received formulization (Greer and

McCalla, 1989) and then developed into concepts of

discrete granularity scales (Hobbs, 1995). We use

Hobbs scales of granularity with the concept of

minimum rough sets to define levels of granularity

from our universe of objects (i.e. our set of original

data sources). Granularity over multiple universes in

rough sets is currently used in decision support and

management science (Słowiński et al., 2014; Sun and

Ma, 2015). We use it here as a formulized approach

to measuring granularity.

Given a universe, U, consisting of a set P of

predicates over a number of objects in O. R is the

relevant subset of predicates from P. So, we can

define objects x and y as indistinguishable if they

meet:

∀



, 



~ ≅



∀ ∈ 











≅







(1)

Two objects are indistinguishable if their values

for every relevant predicate are equal. Expanding on

this, given a set of predicates (or attributes), we can

separate the complete list of objects into sets of

indistinguishable elements or equivalence sets. In the

first definition of aggregation, we define granularity

by the number of attributes used to create the

equivalence sets. In the second definition, each of

these equivalence sets have membership at a discrete

granularity level defined by the number of these

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

indistinguishable sets. So, given a set of attributes we

can determine the granularity level as well as group

data sources for network fusion. In the third definition

of aggregation, we can define granularity by the

number of data sources in the equivalence set and

separate these sets according to granularity.

We created three universes of objects as our

complete set of data sources, using “source”, “tissue”,

“disease state”, “treatment”, “technology”,

“aggregation method”, and “species” as attributes.

The first universe, the overarching universe,

contained all data sources. The second two, diabetes

and control universes, representatively used diabetic

or non-diabetic data sources. In doing so we can see

the effect of experimental condition on information

extraction. We used each combination of every length

of these attributes, defining a set of fused networks

and a defined granularity level. We can use the

enrichment and granularity level to characterize the

relationship between the two, given our defined

universe.

To score information extraction, we use a similar

method as above, measuring the enrichment of

diabetes related terms from Reactome, KEGG, and

OMIM. We define the information extraction score in

two ways. First, we use the proportion of relevant

enrichment terms found over the total number of

relevant enrichment terms. To standardize

enrichment term impact on the score, we also use an

information extraction score which weights the

contribution of an enriched term by the probability of

finding the term, as defined by the calculated

Figure 2: Data Source Sizes – The relative sizes for the data

sources as the number of nodes, number of edges, and

number of clusters are displayed as numbers and

represented as the area of their corresponding circles in

order to show relative size.

probability of finding the term across all used

networks produced by the universe. In this equation,

N is the set of enrichment terms in a given network,

T is the complete set of diabetic enrichment terms,

and P(t) is the probability of finding term t in any

network.

  

∑

1

∈



∑

1

∈

(2)

Then we find the correlation between discrete

levels of granularity and information extraction

scores, utilizing equivalent set derived fused

networks as each point. In total, we obtain six

correlation and p-value pairs for each aggregation

definition from the three universes and the two

scoring techniques. These correlations are applied to

each of the definitions for aggregation granularity:

number of attributes, number of equivalence sets, and

number of contributing data sources.

3 RESULTS

3.1 Protein-protein Interaction

Databases Show Low Structural

Similarity

Structural differences between networks were

examined at the node, edge, and cluster levels. Figure

2 shows the number of biological elements found in

each database, i.e. proteins for the PPI databases, and

transcripts for the microarray series. The overlap

percent of these node sets were calculated by dividing

the intersection of the two sets by their union. As

would be expected the larger databases have low

overlap percent with the smaller databases since their

potential overlap is small. Table 1 shows this overlap

between the PPI networks and the 0.8 power

threshold control microarray network. The larger,

more inclusive networks tend to have higher

similarity, but DIP and MINT, even with a close

number of nodes, had a low overlap.

More so than the overlap between biological

elements, the interactions derived from each data

source showed almost no overlap. The number of

edges in each network seemingly enhanced the

distance in size between data sources. Figure 2 shows

the number of interactions in each network; Table 2

shows the overlap, calculated by taking the

intersection over the union of two interaction sets.

The overlap of clusters from each network was

determined. Only clusters of size five or higher were

used and two clusters needed a 70% member overlap

On the Impact of Granularity in Extracting Knowledge from Bioinformatics Data

as determined by the smallest cluster to be determined

the same. Figure 4 shows the number of clusters in

each network; Table 3 shows the overlap between

these clusters. The BioGRID network had a high

density and clustered into small, yet huge clusters.

Once again, the structural overlap of these networks

is negligible.

Table 1: Node Overlap.

MicroArray DIP Biogrid HPRD IntAct

MINT 0.08 0.22 0.14 0.20 0.20

IntAct 0.18 0.21 0.65 0.50

HPRD 0.16 0.23 0.56

Biogrid 0.19 0.16

DIP 0.09

Table 2: Edge Overlap.

MicroArray DIP Biogrid HPRD IntAct

MINT

0.00 0.02 0.00 0.02 0.04

IntAct

0.00 0.03 0.02 0.06

HPRD

0.00 0.04 0.03

Biogrid

0.00 0.02

DIP

0.00

Table 3: Cluster Overlap.

MicroArray DIP Biogrid HPRD IntAct

MINT

0.00 0.01 0.01 0.01 0.02

IntAct

0.00 0.01 0.08 0.06

HPRD

0.00 0.02 0.06

Biogrid

0.00 0.01

DIP

0.00

Table 4: Potential Enrichment Overlap.

MicroArray DIP Biogrid HPRD IntAct

MINT

0.30 0.52 0.60 0.56 0.57

IntAct

0.74 0.43 0.40 0.40

HPRD

0.71 0.49 0.46

Biogrid

0.72 0.49

DIP

0.54

Figure 3: Full hierarchy of Reactome human pathways. The

entire list of human Reactome pathways and their hierarchy

was graphed so that nodes represent pathways and edges

represent the child-parent relationships in the Reactome

hierarchy. The unique Reactome enrichments for each PPI

network are highlighted in color: HPRD (green), IntAct

(gold), BioGRID (blue), DIP (pink), and MINT (red). These

unique enrichments represent a small proportion of the total

human pathway hierarchy.

Figure 4: Organized hierarchy of Reactome human

pathways - Fig. 3 is restructured to group pathways from

the same branches of the Reactome hierarchy together. The

unique pathway enrichments show low grouping tendency,

and the PPI sources do not demonstrate biological

specificity. Enrichments for each PPI network are

highlighted in color: HPRD (green), IntAct (gold),

BioGRID (blue), DIP (pink), and MINT (red).

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

Figure 5: Relationship between Granularity (Aggregation) Definitions and Information Extraction Scores – For each universe

and each information extraction score type, the correlation and p-value between score type and granularity is shown

(correlation, pvalue). Only significant scores are highlighted and the darker the coloring, the higher the correlation value.

3.2 Reactome Enrichment: Overlap

and Unique Contributions

Although the structural aspects of the networks had

low similarity, the biological properties as

determined through Reactome enrichments had

comparatively high overlap. Table 4 shows the

potential enrichment overlap between data sources.

Here, the intersection is divided by the size of the

smaller enrichment size to highlight the low

availability for unique contributions by the smaller

datasets.

For visualization, we examine only the Reactome

enrichments of each PPI network. The unique

enrichments for individual networks would delineate

any source as specific towards an individual

biological domain. Figure 3 shows the unique

Reactome pathway enrichments for each individual

PPI source in the setting of the entire set of human

Reactome pathways. Although the structural

differences between networks are small, unique

pathway enrichments only make up a small

proportion of the total potential pathway space.

Initially, this indicates that any bias that does exist

towards a specific biological condition or theme, is

weak. However, the impact of these biases can only

be examined by segmenting the Reactome hierarchy

into biologically relevant clusters.

The manual grouping of these pathways into

groups of similar function, as shown in Figure 4, does

not distinguish confident unique themes of biological

extraction. Yet a few sets of unique contributions

show significant grouping within the pathway

hierarchy. HPRD highlights gamma carboxylation;

IntAct highlights GAG protein metabolism; and

BioGRID highlights ER to Golgi transport and

single-nucleotide replacement. However, these

tendencies are insufficient to specify particular

granularity for individual PPI data sources. Rather,

they each demonstrate generic pathway enrichment

across the human pathway hierarchy.

3.3 Data Fusion with PPI Sources

Drowns Microarray Conditional

Differences

We compare the differences between control and

diabetic conditions for the microarray networks, the

PPI networks, and the fusion networks. For the

microarray networks, the false-discovery rate

adjustment diminished the interactions and

enrichment so that there are only three differences

between the control and diabetic conditions. The

power and p-value threshold networks showed ten

pathway differences between the control and diabetic

conditions out of 24 total diabetes related pathways.

After fusion, the network structure for the FDR

adjusted and the thresholded networks were enriched

for nearly every single available pathway. Table 5

shows these enrichments, only showing 0.05 p-value

adjustment. The ten differences from the microarray

networks are not seen in the fusion networks.

3.4 A Significant Interaction between

Granularity and Information

Extraction Exists

The currently defined universes of predicates and

objects has seven relevant attributes; however, we

remove the “source” attribute as it does not produce

equivalence sets greater than one, leaving 60 total sets

of attributes. A total of 57 equivalence sets were

determined across 7 discrete granularity levels for the

On the Impact of Granularity in Extracting Knowledge from Bioinformatics Data

first definition of aggregation granularity, 24 discrete

granularity levels for the second, and 24 for the third.

We found the correlations between the information

extraction scores based on enrichment. Of these, we

found no significant correlations between granularity

and information extraction when using the number of

attributes determining equivalence sets. When using

the number of equivalence sets, half of the correlations

were significant. Finally, when considering the

number of sources as the scale of granularity, each of

the correlations was significant.

Table 5: Pathway Enrichment Network Spectrum.

Micro

Array

Control

0.05 p

Micro

Array

T2D

0.05 p

DIP Bio

GRID

HPRD IntAct MINT Fusion

Control

0.05p

Fusion

T2D

0.05p

Metabolism of lipids

and lipoproteins

PERK regulates gene

expression

Protein processing in

endoplasmic reticulum

Adrenaline,

noradrenaline inhibits

insulin secretion

Calcitonin-like ligand

receptors

Glucagon-like Peptide-

1 (GLP1) regulates

insulin secretion

Signaling by Leptin

Notch signaling

pathway

Wnt signaling pathway

TGF-beta signaling

pathway

Hormone-sensitive

lipase (HSL)-mediated

triacylglycerol

hydrolysis

PPAR signaling

pathway

Cell cycle

p53 signaling pathway

Advanced glycosylation

end-product receptor

signaling

Regulation of insulin

secretion

Unfolded Protein

Response (UPR)

Type II diabetes

mellitus

Diabetes Mellitus,

noninsulin-dependent;

NIDDM

Pancreatic secretion

Maturity onset diabetes

of the young

mTOR signaling

Insulin secretion

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

100

4 DISCUSSION

The cascading importance of high confidence

necessary in medical and biological data puts an

emphasis on reproducibility, stability, and sensitivity

of analytical workflows. Due to the complexity of

biological data analysis, results have a high sensitivity

to small changes in parameters, producing results that

are unstable and difficult to reproduce. This concept is

accentuated and expanded in data fusion, granularity

being just a single potential for sensitivity. The data

fusion approach presented here demonstrates the

sensitivity of these biological data towards the various

levels of granularity along two different dimensions.

4.1 Protein-protein Interaction

Database Differences and

Sensitivity

The structural differences between the protein-

protein interaction networks are byproduct of slight

differences in curation methods. Although each PPI

database retrieves their information from the same

population of publications, the knowledge presented

in the databases is different. These curation

differences which lead to vast node and edge network

differences, also lead to different structural or

clustering differences.

Conversely, although the structure of the

networks is sensitive to the curation methods, the

biological enrichment of the networks is not. The

biological enrichment had a relatively large overlap

between databases. Had the creation of the PPI

databases led to significant biological differences, as

a total, they would have represented a higher

granularity. However, as it stands, the small unique

pathway enrichments are not representative of highly

granular data. To address our hypothesis, H1, the

unique biological extractions of the PPI networks do

not represent any apparent biological themes or

conditions.

4.2 Flooding of High Granular Data

through Data Fusion

The promise of data fusion is a more accurate

depiction of biological reality through the

combination of data sources to compensate for

individual source inadequacies. Yet a key component

to data fusion is within its own definition: the

combination of two or more data elements to create a

novel and meaningful data element. “Meaningful” is

vitally important. A good data fusion result captures

a biologically relevant meaning and treats the data

accordingly.

Hypothesis 2 states that the fusion of low and high

granular data sources will remove experimentally

derived information. The union function utilized here,

innately, favors a low granularity. The PPI databases,

created through the union function, result in

“potential” networks. These networks illustrate the

potential of protein interaction partners and structure,

but may not be specific enough to differentiate

specific biological conditions or pathways, and

ultimately as seen in Table 5 the larger PPI databases

capture the majority of this selection of Type 2

Diabetes enrichments. So, the union between high

granularity data, and low granular data initially

created through the union function, results in a low

granularity data set. As with the PPI networks

covering the potential of interaction partners, this

fusion creates a potential of pathways list, making it

impossible to differentiate between experimental

conditions. I.e. the fusion networks do not show

enrichment dissimilarity. In this case, the

experimental specific differences are those which

differentiate the control tissue from the diabetic

tissue. These differences are flooded and unable to be

extracted after fusion.

4.3 Information Extraction Sensitivity

towards Granularity

The third hypothesis, a granularity scale exists in

which aggregation is associated with information

extraction, was not supported by each of the defined

scales of aggregation. The first scale of aggregation,

number of attributes, did not demonstrate any

relationship with information extraction. When

aggregation was defined as the number of

equivalence sets, half of the correlations were

significant. Explicitly, we find that the diabetes

universe and weighted score using this definition of

granularity had the highest correlation out of all the

tested conditions. The lack of significant correlation

in the control universe indicates that these diabetes

data sources are the reason that significant correlation

was found in the overarching universe. We note that,

biologically, diabetic data sources are likely to have

increased diabetic information extraction; however,

the relationship between granularity and information

extraction is not innately evident. Defining

aggregation by the number of equivalence sets is only

satisfactory under certain data source combinations

and a weighted information extraction score.

The final definition of granularity, as the number

of data sources used in the fused network, had a

On the Impact of Granularity in Extracting Knowledge from Bioinformatics Data

101

complete suite of significant correlations. Yet the

diabetic set of data sources carried a higher bias in

generating a strong relationship between granularity

and information extraction. This result defends the

proposition that more available information innately

present in a data fusion indicates a higher potential for

information extraction after the fusion has taken

place.

By supporting the third hypothesis, we suggest

that sensitivity to granularity contributes to the

confidence in the limits of a data fusion function. Yet

the field of data fusion is diverse and we are not

certain that granularity is important for all data fusion

functions, specifically those which may correct for

abstraction or aggregation among sensor

technologies. Further, granularity is a high-level

uncertainty term and can be defined in various ways

beyond the two dimensions suggested in this study.

An intelligent data fusion must consider the biology,

the technology, and the sensitivity of the function to

initial parameters, including granularity, in order to

obtain confidence in its results.

4.4 Generalizability across Biomedical

Data Sources

When conducting multivariate data analysis,

univariate normality does not guarantee multivariate

normality. In the same way, the sensitivities of

individual biomedical data sources, including the

sensitivity to granularity, must be examined in a

multi-modal perspective. We can only speculate to

the sensitivity of biomedical data sources not

included in this study, but we suggest that while

granularity may not be an issue in an individual data

source, data fusion approaches should check for

sensitivity to granularity.

5 CONCLUSIONS

As the technology associated with biomedical

research continues to advance, larger and more

diverse data sources are becoming available to

researchers. Each data source has its own attributes

that influence the way its data can be used or

integrated with other data. As a result, there is a

growing need for sophisticated ways to effectively

integrate different types of biological data and

improve the outcome of using data mining

algorithms. In this study, we proposed several tests

for characterizing granularity within the integration

of protein-protein interaction and gene expression

data using the network model. The results indicate

that using high aggregation of information provides a

context bias that alters the composition of various

substructures in the network and enhances the

significance of the signals obtained from the

integrated networks, under certain conditions. In

addition, abstraction with union data fusion favors

high abstraction information extraction, flooding

condition-specific results.

This study serves as a case study to highlight the

need to study data integration methods further in the

domain of biomedical informatics and explore

different ways to characterize the impact of

uncertainty variables throughout alternate data

integration methodologies. These characterizations

must also include topological information regarding

substructure changes in order to further classify the

relationships among elements in biological networks.

The underlying principle here is that each network

represents a form of an expert system, the more

relevant data incorporated in the network, the more

knowledgeable the network becomes. Yet the

dependency of the extraction of this knowledge is

dependent on data source variables (including

granularity) which impact the topology. In turn,

proper handling of these variables would allow the

researchers to extract more biologically relevant

signals while limiting the impact of noise that will

always be associated with raw biological data.

Ultimately, the attainment of the more useful

biological networks, dependent on the type and

environment of a network or biological replicate, is

contingent on the ability to successfully integrate data

types through characterization of their sensitivities.

REFERENCES

Agarwal, A. K., Xu, T., Jacob, M. R., Feng, Q., Lorenz, M.

C., Walker, L. A., & Clark, A. M. (2008). Role of heme

in the antifungal activity of the azaoxoaporphine

alkaloid sampangine. Eukaryotic cell, 7(2), 387-400.

Bindea, G., Mlecnik, B., Hackl, H., Charoentong, P.,

Tosolini, M., Kirilovsky, A., Fridman, W., Pages, F.,

Trajanoski, Z., & Galon, J. (2009). ClueGO: a

Cytoscape plug-in to decipher functionally grouped

gene ontology and pathway annotation networks.

Bioinformatics, 25(8), 1091-1093.

Bittner, T., & Smith, B. (2003). A theory of granular

partitions. Foundations of geographic information

science, 7, 124-125.

Bittner, T., Donnelly, M., & Smith, B. (2004, November).

Individuals, universals, collections: On the

foundational relations of ontology. In Proceedings of

the Third Conference on Formal Ontology in

Information Systems (pp. 37-48).

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

102

Bossi, A., & Lehner, B. (2009). Tissue specificity and the

human protein interaction network. Molecular systems

biology, 5(1).

Ceol, A., Aryamontri, A. C., Licata, L., Peluso, D.,

Briganti, L., Perfetto, L., ... & Cesareni, G. (2009).

MINT, the molecular interaction database: 2009

update.Nucleic acids research, gkp983.

Chatr-aryamontri, A., Breitkreutz, B. J., Heinicke, S.,

Boucher, L., Winter, A., Stark, C., Nixon, J., Ramage,

L., … & Tyers, M. (2013). The BioGRID interaction

database: 2013 update. Nucleic acids research, 41(D1),

D816-D823.

Greer, J. E., & McCalla, G. I. (1989, August). A

Computational Framework for Granularity and its

Application to Educational Diagnosis. In IJCAI (pp.

477-482).

Halevy, A., Rajaraman, A., & Ordille, J. (2006, September).

Data integration: the teenage years. In Proceedings of

the 32nd international conference on Very large data

bases (pp. 9-16). VLDB Endowment. Bowman, M.,

Debray, S. K., and Peterson, L. L. 1993. Reasoning

about naming systems. ACM Trans. Program. Lang.

Syst. 15, 5 (Nov. 1993), 795-825.

Hanisch, D., Zien, A., Zimmer, R., & Lengauer, T. (2002).

Co-clustering of biological networks and gene

expression data. Bioinformatics, 18(suppl 1), S145-

S154.

Hobbs, J. R. (1985). Granularity. In In Proceedings of the

Ninth International Joint Conference on Artificial

Intelligence.

Hobbs, J. R. (1995). Sketch of an ontology underlying the

way we talk about the world. International journal of

human-computer studies, 43(5), 819-830.

Ingram, P. J., Stumpf, M. P., & Stark, J. (2006). Network

motifs: structure does not determine function. BMC

genomics, 7(1), 108.

Jiang, P., & Singh, M. (2010). SPICi: a fast clustering

algorithm for large biological networks. Bio

informatics, 26(8), 1105-1111.

Kashani, Z. R., Ahrabian, H., Elahi, E., Nowzari-Dalini, A.,

Ansari, E. S., Asadi, S., ... & Masoudi-Nejad, A.

(2009). Kavosh: a new algorithm for finding network

motifs. BMC bioinformatics, 10(1), 318.

Kerrien, S., Aranda, B., Breuza, L., Bridge, A., Broackes-

Carter, F., Chen, C., ... & Hermjakob, H. (2011). The

IntAct molecular interaction database in 2012.Nucleic

acids research, gkr1088.

Liu, Z., Cao, J., Gao, X., Zhou, Y., Wen, L., Yang, X.,

Xuebiao, Y., Ren, J., & Xue, Y. (2011). CPLA 1.0: an

integrated database of protein lysine

acetylation. Nucleic acids research, 39(suppl 1),

D1029-D1034.

McCalla, G., Greer, J., Barrie, B., & Pospisil, P. (1992).

Granularity hierarchies.Computers & Mathematics

with Applications, 23(2), 363-375.

Medintz, I. L., Vora, G. J., Rahbar, A. M., & Thach, D. C.

(2007). Transcript and proteomic analyses of wild-type

and gpa2 mutant Saccharomyces cerevisiae strains

suggest a role for glycolytic carbon source sensing in

pseudohyphal differentiation. Molecular BioSystems,

3(9), 623-634.

Obayashi, T., & Kinoshita, K. (2009). Rank of correlation

coefficient as a comparable measure for biological

significance of gene coexpression. DNA research,

16(5), 249-260.

Pawlak, Zdzisław (1982). "Rough sets". International

Journal of Parallel Programming 11 (5): 341–

356.doi:10.1007/BF01001956.

Prasad, T. K., Goel, R., Kandasamy, K., Keerthikumar, S.,

Kumar, S., Mathivanan, S., ... & Pandey, A. (2009).

Human protein reference database—2009

update. Nucleic acids research, 37(suppl 1), D767-

D772.

Rector, A., Rogers, J., & Bittner, T. (2006). Granularity,

scale and collectivity: when size does and does not

matter. Journal of biomedical informatics, 39(3), 333-

349.

Rhee, S. Y., Wood, V., Dolinski, K., & Draghici, S. (2008).

Use and misuse of the gene ontology

annotations. Nature Reviews Genetics, 9(7), 509-515.

Salwinski, L., Miller, C. S., Smith, A. J., Pettit, F. K.,

Bowie, J. U., & Eisenberg, D. (2004). The database of

interacting proteins: 2004 update.Nucleic acids

research, 32(suppl 1), D449-D451.

Słowiński, R., Greco, S., & Matarazzo, B. (2014). Rough-

set-based decision support. In Search

Methodologies (pp. 557-609). Springer US.

Sun, B., & Ma, W. (2015). Multigranulation rough set

theory over two universes. Journal of Intelligent &

Fuzzy Systems: Applications in Engineering and

Technology, 28(3), 1251-1269.

Taneera J, Lang S, Sharma A, Fadista J et al. A systems

genetics approach identifies genes and pathways for

type 2 diabetes in human islets. Cell Metab2012 Jul

3;16(1):122-34. PMID: 22768844.

Thorne, T., & Stumpf, M. P. (2007). Generating confidence

intervals on biological networks. BMC bioinfo, 8(1),

467.

Veres, D. V., Gyurkó, D. M., Thaler, B., Szalay, K. Z.,

Fazekas, D., Korcsmáros, T., & Csermely, P. (2014).

ComPPI: a cellular compartment-specific database for

protein–protein interaction network analysis. Nucleic

acids research, gku1007.

Vogt, L., Grobe, P., Quast, B., & Bartolomaeus, T. (2012).

Accommodating ontologies to biological reality–top-

level categories of cumulative-constitutively organized

material entities. PLoS One, 7(1), e30004.

Xu, T., Feng, Q., Jacob, M. R., Avula, B., Mask, M. M.,

Baerson, S. R., ... & Agarwal, A. K. (2011). The marine

sponge-derived polyketide endoperoxide plakortide F

acid mediates its antifungal activity by interfering with

calcium homeostasis. Antimicrobial agents and

chemotherapy, 55(4), 1611-1621.

Zhang, B., & Horvath, S. (2005). A general framework for

weighted gene co-expression network analysis.

Statistical applications in genetics and molecular

biology, 4(1), 1128.

On the Impact of Granularity in Extracting Knowledge from Bioinformatics Data

103