Large Collection of Diverse Gene Set Search Queries Recapitulate Known Protein-Protein Interactions and Gene-Gene Functional Associations

1.9 MB·

by  Avi Ma'ayan

Large Collection of Diverse Gene Set Search
Queries Recapitulate Known Protein-Protein
Interactions and Gene-Gene Functional Associations
Neil R. Clark1,2,3, Avi Ma’ayan1,2,3,∗
1Department of Pharmacology and Systems Therapeutics, Icahn School of Medicine at Mount Sinai,
One Gustave L Levy Place, Box 1215, New York, NY, USA
2Big Data to Knowledge (BD2K) Library of Integrated Network-based Cellular Signatures (LINCS)
6 Data Coordination and Integration Center (DCIC)
1
3Mount Sinai Knowledge Management Center (KMC) for Illuminating the Druggable Genome (IDG)
0
2 ∗Corresponding author: [email protected]

n
a
J

7
Abstract—Popular online enrichment analysis tools from the associated biological theme, e.g., members of cell signaling

fieldofmolecularsystemsbiologyprovideuserswiththeabilityto pathways, or targets of transcription factors. Enrichr [2] is a
]
N submittheirexperimentalresultsasgenesetsforindividualanal- popular online tool we developed to enable users to submit
ysis. Such queries are kept private, and have never before been
M gene sets for this type of analysis. Since its launch in 2013,
considered as a resource for integrative analysis. By harnessing
gene set query submissions from thousands of users, we aim to over900,000genesetshavebeensubmittedtoEnrichrbyover
.
o discover biological knowledge beyond the scope of an individual 25,000 unique users. These gene sets originate from diverse
i study. In this work, we investigated a large collection of gene experimentalplatforms,profilinggenesandproteinsatvarious
b
sets submitted to the tool Enrichr by thousands of users. Based
- regulatory levels of mammalian molecular data.
q on co-occurrence, we constructed a global gene-gene association
We hypothesize that the gene set submitted to Enrichr may
[ network. We interpret this inferred network as providing a
summary of the structure present in this crowdsourced gene set contain patterns and structure that could reveal novel associa-

1 library,andshowthatthisnetworkrecapitulatesknownprotein- tions between genes, gene modules and molecular biological
v protein interactions and functional associations between genes. mechanisms.Thelargesizeofthisdatasetandthediversityof
3 Thisfindingimpliesthatthisnetworkalsoofferspredictivevalue.
the scientific community from which it originates potentially
5 Furthermore, we visualize this gene-gene association network
provide a unique perspective that may reveal new insights.
6 using a new edge-pruning algorithm that retains both the local
1 andglobalstructuresoflarge-scalenetworks.Ourabilitytomake In combinatorial mathematical terms, the nearly 1 million
0 predictions for currently unknown gene associations, that may gene sets that were submitted to Enrichr comprise a family of
1. not be captured by individual researchers and data sources, is a sets,i.e.,acollectionofsubsetsfromthesetofallgenesinthe
demonstrationofthepotentialofharnessingcollectiveknowledge
0 human, mouse and rat genomes. A common approach to the
from users of popular tools in the field of molecular systems
6 analysisofthisformofdataistotreateachsetasanitem-setin
biology.
1
Keywords - gene sets, big data, item set mining analogy to a market basket, and identify frequently occurring
:
v combinationsofitems[5].Analternativehasbeenproposedin
i I. INTRODUCTION whichitem-setsthatarelogicallyrelatedareidentified[6].The
X
Systems approaches to genome-wide molecular data are identificationoflogicalitem-setsisparticularlyrelevantforthe
r
a increasinglyusinggenesets,asopposedtoindividualgenes,as gene sets submitted to Enrichr because this approach takes
thebasicunitsofanalysis.Forexample,inthefieldofmolec- into account rare items, and because biological knowledge is
ular diagnostics, biomarker sets, as opposed to single gene incomplete, logically associated groups of items may only be
biomarkers,areincreasinglybeingapplied[1].Beyondmolec- partially represented in any given basket. These are important
ular diagnostics, genomics, transcriptomics and proteomics properties, as we aim to identify functional modules even
experiments often identify gene sets that are differentially ex- for rarely occurring genes. We expect that due to partial
pressed,orgenesetsthatcontaingeneticvariationsassociated information contained within an individual query, not all
withaphenotype.Usingmethodssuchasenrichmentanalyses members of a relevant functional module will be present in
[2]–[4], common biological functions can illuminate prior any given gene set submitted for analysis with Enrichr. This
knowledge originating from the collection of experimentally partial information is expected to produce high level of false
identified gene sets. positive associations. Due to the expected high false positive
One approach to enrichment analysis is to take a set of rates,adegreeofnoise-filteringmayalsoimprovetheaccuracy
experimentally identified genes and analyze these genes by of such analysis.
comparing them to a library of gene sets with a known Here we provide an interpretable picture of the global
propertiesofalargecollectionofthelistssubmittedtoEnrichr 104
by thousands of users. To accomplish this we employ the Distributionofgenefrequencies
method of logical item-set to extract gene modules. As a con-
sequence, we infer a network of relationships between genes 1000
andvisualizethisnetworkusinganoveledge-pruningmethod
we devised. Next, we examine the degree of compatibility
100
of our findings with annotated gene set libraries such as the nt.
u
Human Gene Atlas or the Gene Ontology (GO), as well as o
C
with known protein-protein interactions. 10
II. RESULTS
1
By the middle of 2015, the Enrichr gene set queries con-
sisted of 172,798 gene sets submitted from 5,114 unique in- 0 200 400 600 800 1000 1200 1400
Enrichrgenefreq
ternetprotocol(IP)addresses.Ouranalysisbeganbyapplying
a number of filters. First, we removed any query that did not Fig. 1. The distribution of the occurrences of unique gene symbols in
correspond to an official gene symbol. Next, we ensured that Enrichrgenesetqueries.
no individual users dominated in their contributions to the
data set by removing all entries from users that contributed
many gene sets (see methods for more details). After this Distributionofquery submissions lengths
strict cleaning criteria, the result was a family of 19,196 gene -3
10
sets, composed of 27,770 genes supplied from 3,308 unique
IP addresses of unique users of Enrichr.
-4
A. Enrichr gene set library metrics nt.10
u
o
Thedistributionofgenefrequenciesissurprisinglycomplex C
(Fig. 1). The distribution approximately decays exponentially -5
10
from frequencies greater than 300; however, the lower fre-
quencies are characterized by a bimodal distribution. The
largemodeatthelowestfrequencieshasanoverrepresentation -6
10
of uncharacterized genes, gene names with no functional 50 100 500 1000
associations or known protein-protein interactions; but even Enrichrlinelength
discounting those genes, this peak remains relatively large.
Fig. 2. The distribution of the size of each gene set query submitted to
The lower frequencies are also characterized by a mode at
Enrichr.
around200occurrences.Thefrequencyofoccurrenceofgenes
in submissions to Enrichr has a characteristic frequency that
decays exponentially, as opposed to being characterized by a In order to visualize the network of gene associations, we
powerlaw.Giventhepredominanceofscale-freedistributions, employ an edge-pruning algorithm that results in an inter-
it is perhaps surprising that there is a definite scale for pretable representation of the global network structure while
frequencies of gene occurrence in this data set. preserving local features of the network topology (Fig. 3a). A
The distribution of the lengths of gene sets submitted to selection of the largest local structures, preserved by the edge
Enrichr approximately fits a power-law, with a notable peak pruning algorithm, is shown in Fig. 3b. One potential use of
at about 200 and 400 genes (Fig. 2). The distribution of the thisnetworkistomakepredictionsofnovelassociationsbased
number of submissions to Enrichr per user IP address also onpriorknowledge.Asanexampleofthis,wehighlightgenes
follows a power-law, suggesting that most users submit only known to be associated with Adult Onset Diabetes (KEGG)
a few lists, but there is a substantial body of heavy users. (Fig. 3c). Illustrating local network structures that include at
least one of these genes suggests local structures as candidate
B. Gene-gene and gene-set/gene-set network inference
novel predictions for genes likely associated with the adult-
TheEnrichrcollectionofgenesetqueriescanbetransposed, onset diabetes pathway.
such that for each gene there is an associated set of Enrichr Oneofthelocalstructurescontainstheadult-onsetdiabetes
submission queries. To explore the structure in this data, Genes MAFA and NEUROG3 (KEGG) along with two other
and also to potentially form the basis for an informative genes,UTS2RandFLJ45717,thatarenotidentifiedinKEGG
decomposition of the data, we can infer networks of similar- as being associated with diabetes. However, UTS2 and its
ities between gene pairs and, in addition, networks of Enricr receptor UTS2R have been reported [7] to be involved in
queries. We employ the methods developed in [6] to infer glucose metabolism and insulin resistance, which lead to the
noise-filtered networks of associations between these entities. developmentoftype-2diabetesinhumans.Thereisnoknown
b
a
c MNX1
NR5A2 MAFA
HNF4A FOXA3
INS BHLHB8
ONECUT1
FOXA2
HNF4G
NEUROG3
Fig.3. Theedge-prunednetworkofgeneassociationsinferredfromEnrichrqueries.(a)Theglobalviewofthenetwork.(b)Aselectionofthelargestlocal
networkstructures.(c)Asanexampleofonepossibleuseofthisnetwork,wehighlightgenesknowntobeassociatedwithadult-onsetdiabetes(KEGG)and
illustratethelocalnetworkstructurethatincludesatleastoneofthesegenes.
association of FLJ45717 with diabetes; however, as this gene D. Examiningthegenenetworkforrecoveryofknownprotein-
forms the center of a star graph local structure, we consider protein interactions
it a novel candidate for involvement in adult-onset diabetes.
Next, we asked whether the gene-gene association network
We can also apply the same approach to the network of
created from the Enrichr queries can be used to predict
associations between gene set queries submitted to Enrichr
physicalproteininteractions.WeusedthePSICQUICdatabase
(Fig. 4). This network appears to show a degree of clustering
of protein-protein interactions (PPIs) and examined the statis-
by user; this perhaps reflects individual users special research
tical significance of the overlap between the PSICQUIC PPI
interests, or a common method of data acquisition.
networkandthegene-genenetworkinferredfromEnrichrsub-
missions.Afterapplyingathresholdof0.3forthesimilarityof
anormalizedpoint-wisemutualinformationforthegene-gene
association network, two binary matrices remained.
C. Comparing the clustering of Enrichr networks to random
In the first test of the significance of the similarity between
scale-free networks
these two binary matrices, we count the number of non-
zero entries in each matrix and the number of elements that
The clustering coefficients of the gene and gene-set net- are nonzero in both matrices. Based on a null distribution
works are 0.43 and 0.17, respectively. In order to gauge the whereby the matrices are randomly permuted, we can use the
significance of this, we compared these clustering coefficients hypergeometric distribution to quantify the significance of the
to the typical clustering coefficients for a random scale-free overlap between the two matrices.
network generated by the Barabasi-Albert random scale-free There are n = 13,586 genes that are in both the Enrichr
1
graphmodel.Artificialshufflednetworksofasimilarsizeand queries network and the PSICQUIC PPI network. With a
degree distribution generated by the Barabasi-Albert model threshold on the normalized point-wise mutual information of
have a typical clustering coefficient of 0.0026 and 0.0054 0.05,therearen =99,758edgesintheEnrichrnetworkand
1
respectively. Hence, we see that the Enrichr queries gene, and n =156,388 known PPIs. There are n =2,763 edges that
2 k
gene-set,networkshaveaclusteringcoefficientwhichisorders are present in both the Enrichr network and the PSICQUIC
of magnitude greater than would be expected if the networks PPI network. In this regime, the Poisson approximation to the
were randomly scale-free by the Barabasi-Albert model. hypergeometricdistributionisapplicable.Withatotalnumber
(a)
(b)
Fig.4. Theedge-prunednetworkofassociationsbetweenquerysubmissionstoEnrichr.(a)Theglobalviewofthenetwork.(b)Aselectionofthelargest
localstructures.(c)ThegenesetssubmittedbythefiveEnrichruserswiththelargestnumberofsubmissionsarehighlighted,eachwithadifferentcolor.
of possible edges given by n (1−n ), the rate parameter for
g g
the Poisson distribution is given by:
n n
1 2 =169.1 (1)
n (1−n )
g g
Under the null hypothesis that edges are randomly assigned,
weexpectameannumberofsharededgestobe169.1withthe
standard deviation of approximately 13.0. Consequently, the
observedvalueofn =2,763isabout200standarddeviations
k
greater than the expected value. Therefore, under the null
hypothesis, the number of edges present in both the Enrichr
gene network and the PSICQUIC PPI network is extremely
statistically significant.
It is conceivable that this result is due to users submitting
gene set queries containing proteins known to have many
interactors, and hence the null distribution we used for the
calculation would be inappropriate and the result invalid.
One way to account for this is to condition the test on the
frequency of each gene. To do this, we perform a separate
hypergeometric test for each gene, and correct for multiple
hypothesis testing. There are 899 proteins that have at least
one interaction that overlaps with the Enrichr network; when
corrected for multiple hypotheses testing, with a False Dis- Fig.5. Thepartofthenetworkofthegene-geneassociationsinferredfrom
covery Rate of 2 Enrichr that is also supported by known protein-protein interactions from
PSICQUIC.
Another possible explanation for the apparently significant
overlap between edges in the Enrichr gene network and the
PSICQUIC PPI network is that there are a significant number
of user queries containing genes with known PPIs. To address
this, we recovered the Enrichr lists that contribute to the
prediction of at least one PPI and counted the number of
the significant overlap in the edges in the gene network
proteins in each query list for which there is another protein
inferred from Enrichr submissions is predictive of known and
in the query with which it interacts. If a user submitted a
potentially novel protein-protein interactions.
list of proteins with known protein interactions, then the ratio
of the counted proteins to the length of the query will be
large. We note that the actual ratio is less than 1%, which We show the network of edges between genes that are
suggests that users are not submitting gene sets with known present in both the Enrichr-inferred gene network and the
PPIs to Enrichr to any significant degree. This suggests that PSICQUIC PPI network (Fig. 5).
TABLEI
OMIMExpanded
VTiFru PsMWIMNTs GENESETLIBRARIESUSEDTOASSESSRECOVERYOFFUNCTIONAL
micTroFRPNPAIss ASSOCIATIONSBYTHENETWORKINFERREDFROMTHEENRICHR
BioCartapathwKaEyAs QUERIES.
MGIMPtop3
MGIMPtop4
NURSAIP-MS
CORUM GeneSetLibrary #ofSets MeanSetSize
GeneSigDB
GeneOntologyCC BioCartapathways 249 18
GeneOntologyBP
PPIHubProteins CancerCellLineEncyclopedia 967 176
NCI60
HMDBMetabolites ChEA 240 1456
Structural Domains
GeneOntologyMF Chromosomelocation 386 85
Chromosomelocation
KEWGikGiPpaatthhwwaayyss CORUM 1673 5
OMIMdiseaseCgeCnLeEs GeneOntologyBiologicalProcess 941 78
ReactomepathCwhaEyAs GeneOntologyCellularComponent 205 172
HMuomuasenGGeenneeAAttllaass GeneOntologyMolecularFunction 402 122
0.0 0.1 0.2 0.3 0.4 0.5 GeneSigDB 2139 127
Fractionofgenesetssignificantlyclustered(FDR1%) GenomeBrowserPWMs 615 275
HMDBMetabolites 3906 47
Fig. 6. The proportion of gene sets in each library that are significantly
HumanGeneAtlas 84 450
clustered in the gene-gene association network inferred from queries to
KEA 474 37
Enrichr.
KEGGpathways 200 48
MGIMPtop3 71 717
MGIMPtop4 476 202
E. Evaluating the recapitulation of prior knowledge about microRNAs 222 155
MouseGeneAtlas 96 660
genefunctionswiththegene-geneassociationnetworkinferred
NCI60 93 343
from Enrichr queries NURSA-IP-MS 1796 158
OMIMdiseasegenes 90 25
We took a collection of 28 annotated gene set libraries that OMIMExpanded 187 89
represent prior knowledge of associations between genes and Pfam-InterPro-domains 311 35
PPIHubProteins 385 247
their functions from a range of biological themes, including
Reactomepathways 78 73
GeneOntology,mousephenotypes,proteincomplexes,histone TFPPIs 290 79
modificationsandDNAbinding.Weaddressedthequestionof VirusMINT 85 15
WikiPathwayspathways 199 39
the extent to which the information in these gene set libraries
isrecapitulatedbythegene-geneassociationnetworksinferred
by Enrichr queries.
sets in the library are recoverable by the Enrichr queries’
In order to test this for a single gene set, we mapped the
gene-gene association network. The libraries containing data
genes in the set to nodes in the Enrichr queries gene-gene
from genome-wide gene expression profiling have the most
association network, and then calculated the mean point-wise
overlap,followedbyChEA,whichhaslistsofgenesassociated
mutual information between each pair of genes. In order to
with transcription factors from ChIP-seq profiling. Next are
assess the significance of this collective measure of similarity,
pathwaylibrariessuchasreactome,KEGGandWikiPathways.
we numerically calculated a null distribution by randomly
This rank order of libraries likely reflects the nature of
choosing a set of nodes in the Enrichr queries’ network with
most submitted lists to analysis with Enrichr. These query
the same cardinality as the gene set in question. The random
submissions are likely dominated by differentially expressed
choice was weighted by the overall frequency of the gene in
genes from genome-wide mRNA profiling. The recovery of
thegenesetlibraryfromwhichitoriginates.Thisisequivalent
annotated pathways suggests that the inferred gene-gene as-
to generating a null distribution based on random permutation
sociation network likely contains new and more complete
ofthegenesetlibraryrepeatedmanytimes.Bycomparingthe
pathway knowledge.
actualcollectivesimilaritytothenulldistribution,wecalculate
a significance p value for the collective similarity of the gene III. DISCUSSIONANDCONCLUSIONS
set in the Enrichr queries’ gene-gene association network. A In this study, we analyzed for the first time, a large collec-
small p value indicates that the members of the gene set in tion of gene set queries submitted by the many users of the
question are significantly more similar to each other in the popular enrichment analysis tool Enrichr. We examined the
network inferred by Enrichr queries than would be expected distribution of occurrence of genes, the length of submitted
if the gene set library from which the set was originated was gene sets, and the distribution of submissions by individual
randomlypermuted.Weinterpretasmallpvalueasindicating users. We show that the distribution of gene occurrence in
that the network inffered by theEnrichr queries has recovered queries is complex, with some genes appearing often in
the gene set in question. The list of all gene set libraries submitted lists. The length of gene set queries peaks at 200-
examined (Table I). 400genesperquery,andthedistributionofsubmittersfollows
After correcting for multiple hypotheses testing and setting apowerlaw.Byconstructingagene-geneassociationnetwork
afalsediscoveryratethresholdof1%,wecountedthenumber of co-occurrence, we were able to show that such a network
of gene sets in each library that are significantly recovered in captures known protein-protein interactions and functional
the Enrichr queries’ gene-gene network (Fig. 6). associationsmuchmorefrequentlythanbychance.Thismeans
We observed that for some libraries, over 30% of the gene that the collective data from thousands of queries potentially
holds new knowledge about the local and global structure B. Network Inference
of the human functional interactome. As more submissions
We employ the methods described in [6] to reconstruct
accumulate,itisexpectedthatsuchpredictionswillberefined
networks from the filtered gene sets.
and improved. While experimental validation of these predic-
Themostbasicapproachtoconstructanetworkfromalarge
tions is outside the scope of this present work, the gene-gene
collection of gene sets is to define a distance matrix between
associationnetworkresourcewedevelopedherecanbeusedto
the gene sets and examine the resulting distance matrix. The
prioritize predictions for general and specific hypotheses that
mostobvioussimilaritymeasurebetweentwogenesetsAand
can be tested experimentally and systematically drive rational
B is the Jaccard index:
research explorations. The reuse of privately submitted gene
|A∩B|
sets by the Enrichr users should be handled with caution. The J(A,B)= (2)
|A∪B|
analysisthatweconductedhereandthegene-geneassociation
networks we constructed keep the identities of users private. Let the set of all elements be V, and the family of sets be M,
Furthermore, we do not provide the gene sets used for our then:
analysis publicly. M =(cid:26)m(n) =(cid:110)m(n)(cid:111)Ln ⊂V(cid:27) (3)
l
l=1
IV. METHODS and the co-occurrence counts of pairs of genes is given by:
A. Data Preprocessing (cid:88)N
φ(α,β)= δ(α∈m(n))δ(β ∈m(n)) (4)
As of June 25th 2015, 172,798 queries of gene sets were
n=1
submitted to Enrichr from 5,114 unique IP addresses. In
where δ(bool) is a Dirac delta function defined such that:
order to analyze the global structure of these gene sets,
(cid:40)
preprocessing was necessary because the raw data include 1 ifbool=True,
δ(bool)=
multiple instances of the example data set supplied by the
0 Otherwise
Enrichrwebsitefordemonstrationpurposes,specialcharacters
and strings that cannot be interpreted as gene lists caused by The marginal counts are then defined in terms of the co-
erroneous input from users, and large collections of queries occurrence counts as:
from individual users who utilized the Enrichr API. These (cid:88)
φ(α)= φ(α,β) (5)
entries were removed so that the resulting dataset is repre-
sentative of the entire community of Enrichr users and is not β∈G,α(cid:54)=β
dominated by any individual user or small subset thereof. whichisthenumberofallco-occurrences.Finally,thetotal
The first filter was applied to retain only word strings that number of co-occurrences is given by:
are members of a reference set of 39920 standard human, 1 (cid:88) 1 (cid:88) (cid:88)
mouse and rat gene names. The resulting sets of genes φ0 = 2 φ(α)= 2 φ(α,β) (6)
corresponding to each user input list was then subjected to α∈V α∈V β∈V
a sequence of subsequent filters. Then the co-occurrence probability is given by:
Use of the Enrichr API enables programmatic access to
φ(α,β)
Enrichr.Thisresultedinaminorityofuserssubmittingalarge P(α,β)= (7)
φ
number of gene lists. In order for the data set to be balanced 0
and representative of the entire community of Enrichr users, and the marginal probability is given by:
and not dominated by a small minority of users, we filtered
P(α)=φ(α)/phi (8)
out gene lists associated with IP addresses from which large 0
numbers of lists were received. The Jaccard index can now be defined in these terms as:
The results of 11,579 differential expression analyses were
P(α,β)
received from the GEO2Enrichr Chrome Extension, which ψ = (9)
jac P(α)+P(β)−P(α,β))
submits three gene sets for every differential expression com-
putationforanalysiswithEnrichr:upregulated,downregulated When the collection of gene set queries is transposed such
andcombinedgenesets,respectively.Onlythecombinedgene that the genes are the labels and the queries are the members
sets were retained in this analysis. of sets, the Jaccard index can be applied.
Gene sets that contained more than 2,000 genes were We can also use the following distance measures:
rejected on the basis that they were non-specific and incurred • Cosine Distance:
(cid:112)
undue computational expense in subsequent analysis. ψ =P(α,β)/ P(α)P(β)
cos
Finally,allinstancesofthedemonstrationgenesetfromthe • Point-Wise Mutual Information:
Enrichr website were removed. The result was 19,196 gene ψ =max{0,Log(P(α,β)/P(α)P(β))}
pmi
sets composed of 27,770 genes supplied from 3,308 unique • Normalized Point-Wise Mutual Information:
IP addresses. ψ =−ψ /(Log(P(α,β)))
nmi pmi
We apply a noise filter by iteratively applying a threshold for ACKNOWLEDGMENT
similarity: the pair counts for gene pairs that do not pass
We would like to thank Dr. Kathleen M. Jagodnik for pro-
the threshold are set to zero, and the similarity matrix is
viding useful comments and copyediting. Funding: This work
recalculated.
wassupportedinpartbygrantsfromtheNIH:R01GM098316,
Throughout the rest of this analysis, we only use the U54HG008230 and U54CA189201 to AM.
normalized point-wise mutual information as a measure of
similarity between genes for the Enrichr queries data. REFERENCES
[1] X.Wang,E.Dalkic,M.Wu,andC.Chan,“Genemodulelevelanalysis:
C. Network visualization identificationtonetworksanddynamics,”Currentopinioninbiotechnol-
ogy,vol.19,no.5,pp.482–491,2008.
Becausetheresultantnetworksaredenselyconnected,direct
[2] E.Y.Chen,C.M.Tan,Y.Kou,Q.Duan,Z.Wang,G.V.Meirelles,N.R.
visualization of the graph is not interpretable. Edge pruning Clark,andA.Ma?ayan,“Enrichr:interactiveandcollaborativehtml5gene
has been proposed for the simplification of graphs to aid in listenrichmentanalysistool,”BMCbioinformatics,vol.14,no.1,p.128,
2013.
generating interpretable visualizations of their structure [8].
[3] D. W. Huang, B. T. Sherman, and R. A. Lempicki, “Bioinformatics
We find that such edge pruning algorithms, applied to very enrichmenttools:pathstowardthecomprehensivefunctionalanalysisof
large graphs, are able to represent the global structure at the largegenelists,”Nucleicacidsresearch,vol.37,no.1,pp.1–13,2009.
[4] A. Subramanian, P. Tamayo, V. K. Mootha, S. Mukherjee, B. L. Ebert,
expense of the local structure. To potentially improve this, we
M.A.Gillette,A.Paulovich,S.L.Pomeroy,T.R.Golub,E.S.Lander
employed an edge pruning approach that is able to capture et al., “Gene set enrichment analysis: a knowledge-based approach
both global and local structure in the graph. for interpreting genome-wide expression profiles,” Proceedings of the
NationalAcademyofSciencesoftheUnitedStatesofAmerica,vol.102,
In the initial step, a set of local structures is derived by
no.43,pp.15545–15550,2005.
varying the threshold for similarity and identifying connected [5] S. Brin, R. Motwani, J. D. Ullman, and S. Tsur, “Dynamic itemset
graph components at the point at which they disconnect from countingandimplicationrulesformarketbasketdata,”inACMSIGMOD
Record,vol.26,no.2. ACM,1997,pp.255–264.
the giant component. Such point is defined as any connected
[6] S.Kumar,V.Chandrashekar,andC.Jawahar,“Logicalitemsetmining,”
component that contains more than 10% of all nodes or has in Data Mining Workshops (ICDMW), 2012 IEEE 12th International
an absolute number of nodes greater than 100 nodes; this Conferenceon. IEEE,2012,pp.603–610.
[7] Z. Jiang, J. J. Michal, D. J. Tobey, Z. Wang, M. D. MacNeil, and
number is chosen based on the visualizability properties of
N. S. Magnuson, “Comparative understanding of uts2 and uts2r genes
the resulting local structures. for their involvement in type 2 diabetes mellitus,” International journal
The local structures represent local regions of the network ofbiologicalsciences,vol.4,no.2,p.96,2008.
[8] F. Zhou, S. Malher, and H. Toivonen, “Network simplification with
that are distinguishable from the rest of the network at an ap-
minimallossofconnectivity,”inDataMining(ICDM),2010IEEE10th
propriatescaleofsimilarity.Assuch,theydepictthesimilarity InternationalConferenceon. IEEE,2010,pp.659–668.
relationshipsbetweenthenodesandtheirneighborhoodinthe
network.
Once the local structures have been identified, we employ
a naive edge-pruning algorithm. We do not prune edges that
are members of the local structures, thereby preserving them
in the final simplified network. The number of edges that are
available to be pruned is equal to the difference between the
total number of edges in the original network, n , and the
e
numberofedgesthatarepartofthelocalstructures,n .The
e;ls
pruning process preserves the connectivity of the network, so
after the maximum degree of pruning, the minimum number
of edges remaining must be equal to:
n +n (10)
ls e;ls
Accordingly, the maximum number of edges that can be
pruned while preserving the connectivity of the network is:
n −n −n (11)
e e;ls ls
Theiterativeprocessbywhichedgesareprunedisasfollows:
1) Set a counter i=0
2) Identify the edge with the lowest similarity measure
3) If removing the edge does not cut the network, then
remove it.
4) i→i+1
5) Ifi<γ(n −n −n )thenreturntostep1;otherwise,
e e;ls ls
stop.