Discovery of biological networks from diverse functional genomic data.

Abstract

We have developed a general probabilistic system for query-based discovery of pathway-specific networks through integration of diverse genome-wide data. This framework was validated by accurately recovering known networks for 31 biological processes in Saccharomyces cerevisiae and experimentally verifying predictions for the process of chromosomal segregation. Our system, bioPIXIE, a public, comprehensive system for integration, analysis, and visualization of biological network predictions for S. cerevisiae, is freely accessible over the worldwide web.

Full text

Page 1

Genome Biology 2005, 6:R114
comment
reviews
reports
deposited research
refereed research
interactions
information
Open Access
2005Myers et al.Volume 6, Issue 13, Article R114
Method
Discovery of biological networks from diverse functional genomic
data
Chad L Myers*†, Drew Robson‡, Adam Wible*, Matthew A Hibbs*†,
Camelia Chiriac†, Chandra L Theesfeld§, Kara Dolinski† and
Olga G Troyanskaya*†
Addresses: *Department of Computer Science, Princeton University, 35 Olden Street, Princeton, NJ 08544, USA. †Lewis-Sigler Institute for
Integrative Genomics, Carl Icahn Laboratory, Princeton University, Princeton, NJ 08544, USA. ‡Department of Mathematics, Princeton
University, Washington Road, Princeton, NJ 08540, USA. §Department of Genetics, School of Medicine, Mailstop-S120, Stanford University,
Stanford, CA 94305-5120, USA.
Correspondence: Olga G Troyanskaya. E-mail: ogt@cs.princeton.edu
© 2005 Myers et al; licensee BioMed Central Ltd.
This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Biological networks discovery<p>BioPIXIE is a probabilistic system for query-based discovery of pathway-specific networks through integration of diverse genome-wide data.</p>
Abstract
We have developed a general probabilistic system for query-based discovery of pathway-specific
networks through integration of diverse genome-wide data. This framework was validated by
accurately recovering known networks for 31 biological processes in Saccharomyces cerevisiae and
experimentally verifying predictions for the process of chromosomal segregation. Our system,
bioPIXIE, a public, comprehensive system for integration, analysis, and visualization of biological
network predictions for S. cerevisiae, is freely accessible over the worldwide web.
Background
Understanding biological networks on a whole-genome scale
is a key challenge in modern systems biology. Broad availabil-
ity of diverse functional genomic data from protein-protein
interaction, gene expression, localization, and regulation
studies should enable fast and accurate generation of network
models through computational prediction and experimental
validation. Reliability of experimental results varies among
data sets and technologies, however, and these data generally
provide only pair-wise evidence for biological relationships
between genes or proteins. Most cellular mechanisms, on the
other hand, involve groups of genes or gene products that
behave in a coordinated way to perform a specific biological
process. We will refer to such groups of functionally related
genes as process-specific networks. Although a wide variety of
functional genomic data is available, and much has been
learned from them, we are far from exploiting the full poten-
tial of these data for discovering such process-specific net-
works. There are several reasons for this: lack of accessibility
to data and methods to analyze them, barriers to incorporat-
ing expert knowledge in the network discovery process, and
noise and heterogeneity in high-throughput gene data.
The first problem is simply the lack of accessibility of both the
data and analysis methods. Even when data are publicly avail-
able, results are often buried in large files, and computational
methods developed to analyze them are often not available in
forms that the typical biologist can use. Thus, experimental
researchers are unable to identify interesting results from
computational studies that are worth verifying. Instead, most
biologists are limited to what the authors of such studies
deem important or interesting enough to highlight in the
written publication. Our ability to effectively utilize genomic
data for process-specific network discovery has thus been
Published: 19 December 2005
Genome Biology 2005, 6:R114 (doi:10.1186/gb-2005-6-13-r114)
Received: 1 July 2005
Revised: 31 August 2005
Accepted: 21 November 2005
The electronic version of this article is the complete one and can be
found online at http://genomebiology.com/2005/6/13/R114

Page 2

R114.2 Genome Biology 2005, Volume 6, Issue 13, Article R114 Myers et al.
http://genomebiology.com/2005/6/13/R114
Genome Biology 2005, 6:R114
hampered by the lack of effective interfaces to both the data
and the relevant analysis methods.
The second challenge is to allow biology researchers to inte-
grate their biological knowledge in analysis. When biologists
inquire about particular biological processes, they bring with
them existing knowledge that can and should be used to gen-
erate the most sensitive and precise hypotheses possible.
Such information is hard to extract automatically, and effec-
tively incorporating biological expert knowledge is of course
closely linked to the accessibility challenge noted above. Most
previous methods for process-specific network prediction
have not allowed biologists to use their previous knowledge in
their area of interest to target the analysis process. Biological
research demands convenient and accessible systems that
leverage existing knowledge to direct and facilitate discovery.
The third challenge in constructing accurate process-specific
networks from diverse genomic data lies in the heterogeneity
and high noise levels in large-scale data sets. High-through-
put data by nature are often noisy and simple combinations of
results from different types of experiments (for example, con-
clusions of genome-scale two-hybrid experiments and micro-
array studies) are of limited effectiveness because they
sacrifice either sensitivity or specificity.
Recent applications of probabilistic data integration to the
related but simpler problem of predicting protein function
from diverse genomic data have demonstrated that inte-
grated analysis of heterogeneous sources provides a substan-
tial increase in prediction accuracy. Much of the work in
function prediction focuses on fusing information from mul-
tiple heterogeneous sources for pairs of proteins to make
more reliable statements about pair-wise functional relation-
ships. Bayesian networks [1,2] and variations of this approach
[3-5] have been applied successfully to construct 'functional
linkage maps' whose connecting edges represent probabilistic
support for a functional relationship between the adjacent
proteins. Protein functions are then inferred through 'guilt by
association' with surrounding nodes of known function. Sev-
eral studies have formalized this 'guilt by association'
approach by using Markov Random Field models to propa-
gate known functional annotations through confidence-
weighted edges [6-8].
Despite much investigation into heterogeneous data integra-
tion for the purpose of function prediction, there have been
only limited attempts to use confidence-weighted linkage
maps from integrated data to address the more biologically
significant problem of how to group functionally related pro-
teins together into process-specific networks. These network-
level questions are distinctly different from function predic-
tion problems and require new methodology for general data
integration and network discovery. Previous work in identify-
ing groups of genes involved in specific biological pathways
from interaction networks has focused on mainly binary
interactions, which are prone to false positives and inade-
quate coverage when only limited types of genomic evidence
are used. For instance, two studies [9,10] describe
approaches for finding highly connected subgraphs in binary
interaction graphs from high-throughput experiments. They
found that highly connected groups in these graphs often cor-
respond to protein complexes or biological processes.
Another study [11] introduced the notion of modular decom-
position of protein-protein interaction networks to make
inferences about pathways. While these approaches have
demonstrated the promise of using protein-protein interac-
tion networks for recognizing groups of proteins involved in
specific processes, they are constrained by their reliance on
limited types of interaction data and their use of binary,
rather than probabilistic networks. A recent study extended
these approaches to a weighted interaction network and used
graph clustering analysis to detect coordinated functional
modules [12]. A common theme among many of these studies
is their unsupervised approach to network detection. Incor-
porating expert knowledge in the search process, however,
can dramatically improve both the specificity and sensitivity
of process-specific network discovery from protein-protein
interaction data.
To our knowledge, the only existing work that leverages
expert knowledge in constructing biological networks or pro-
tein complexes from integrated data is a network reliability
approach to protein complex recovery [13] and a greedy
search algorithm applied to a confidence-weighted protein-
protein interaction network [14]. The former was specifically
targeted towards protein complexes, while we focus on the
more general problem of discovering not just physically inter-
acting sets of proteins, but functional or process-specific net-
works. The latter algorithm, proposed by Bader [14],
leveraged both physical and genetic interaction data with the
goal of extracting more general protein networks. Distinc-
tions between Bader's and our approach are that we integrate
functional genomic data in a Bayesian framework that allows
a probabilistic, rather than heuristic, graph search. This prob-
abilistic search incorporates both direct and indirect protein-
protein links while integrating a wider variety of data (for
example, microarray expression, co-localization). Further-
more, we are the first to our knowledge to develop an interac-
tive, web-accessible system that both facilitates discovery of
novel biological networks and allows exploratory analysis of
the underlying genomic data that support these predictions.
To address these challenges to discovering process-specific
networks from functional genomic data, we have created a
publicly available system called bioPIXIE (biological Process
Inference from eXperimental Interaction Evidence). The sys-
tem allows users to enter a set of proteins and then uses a
novel probabilistic graph search algorithm on a protein-pro-
tein linkage map derived from diverse genomic data to pre-
dict the surrounding process-specific network for the local
neighborhood of interest. Most importantly, the system

Page 3

http://genomebiology.com/2005/6/13/R114
Genome Biology 2005, Volume 6, Issue 13, Article R114 Myers et al. R114.3
comment
reviews
reports
refereed research
deposited research
interactions
information
Genome Biology 2005, 6:R114
includes a convenient interface for dynamic visualization of
the resulting predictions and provides analysis of their func-
tional coherence. We have completed an extensive evaluation
of our method against known pathways as well as experimen-
tally verified a subset of predictions made by our system.
Results
Evaluation of the method on known biological
networks
Our system achieves accurate network prediction by effec-
tively integrating diverse data sets and probabilistically iden-
tifying new components of process-specific networks given
only one or a few known members. We evaluated the ability of
our approach to recover known process-specific networks
given initial query sets by using a collection of well-annotated
functional groups, including KEGG pathways, sets of biologi-
cal process GO terms, and MIPS protein complexes. We
restricted our evaluation to groups of 15 to 250 total proteins
in which at least half of the member proteins had one type of
evidence linking them with another member protein. We
identified 31 such groups from the set of KEGG pathways,
MIPS protein complexes, and GO terms (see Additional data
file 2 and supplemental Table S1 in [15]). We evaluated the
performance of our method on each group by sampling 100
random query sets consisting of 10 proteins each from the
pathway or complex of interest, applying our data integration
and search algorithm, and analyzing the returned set of pro-
teins for consistency with the remaining proteins in the
group.
The advantage of using bioPIXIE to integrate multiple types
of genomic data is illustrated in Figure 1a-c for three diverse
KEGG pathways (graphs for all 31 processes are available in
supplemental Figure S2 in [15]). bioPIXIE dramatically and
consistently improves the number of network components
recovered over any of the individual types of evidence. For
example, for KEGG cell cycle proteins (Figure 1a), given a ran-
dom 10-protein query set, we identified an average of 42 of
the remaining 77 proteins using integrated data, whereas only
25 were identified by either physical or genetic evidence, and
only 18 by microarray evidence alone. Different evidence
types have varying degrees of relevance for different path-
ways - microarray correlation is very informative for ribos-
ome proteins (Figure 1b) whereas physical interactions are
more informative for proteins involved in ATP synthesis (Fig-
ure 1c).
This advantage of integrating diverse data types is confirmed
in a more comprehensive evaluation of bioPIXIE's perform-
ance, where we averaged results over the entire set of 31 proc-
esses and complexes described above. Figure 1d compares the
precision-recall characteristics of our network identification
method using Bayesian integrated data versus using individ-
ual evidence types. Given only 10 query genes, the integrated
version recovered 50% of the remaining members at a preci-
sion of 30% whereas the method applied to independent sub-
sets achieved only 15% (physical association), 10% (genetic
association), and 3% (microarray correlation) precision at the
same recall (Figure 1d). Thus, combining data from multiple
sources clearly improves network recovery.
One might expect that due to the relative sparseness of cur-
rent functional genomic data, simple combinations of these
sources followed by a straightforward search would be suffi-
cient for precise network recovery. However, such combina-
tions are substantially less effective than our approach, as
shown in Figure 1e, which plots the average precision-recall
characteristics of two such approaches to integration and
recovery. The first approach ('Binary recovery') uses all avail-
able evidence, but only as a binary 'yes' or 'no', depending on
whether evidence of any type is present for a particular pro-
tein pair. Given a query, connected proteins are then added in
an arbitrary order. The second approach ('Counting-based
recovery') also uses all available evidence but counts observed
evidence for each pair such that overlaps between multiple
sources of evidence receive higher weights. Proteins are then
added in order of weight for network recovery. Neither of
these simpler approaches achieves accuracy similar to that of
our method. In fact, the counting-based approach yields a 4-
fold lower prediction precision than our approach and the
binary approach results in a 10-fold lower prediction preci-
sion at 50% recall.
In addition to these two naive methods, we have also com-
pared our system to two previously published methods for
query-based protein complex discovery, SEEDY [13] and
Complexpander [14]. bioPIXIE's performance is superior to
both existing methods; it achieves an average of 30% preci-
sion at 50% recall while SEEDY yields 12% and Complex-
pander 7% at 50% recall (Figure 1f). Furthermore, calculating
the average area under the precision-recall curve (AUC) for
each pathway individually, we find that the average bioPIXIE
AUC exceeds the average SEEDY AUC by more than one
standard deviation for 22 of the 31 groups, while SEEDY out-
performs only bioPIXIE for only 1 of the 31 groups (Addi-
tional data file 3 and supplemental Figure S4 in [15]).
Similarly bioPIXIE outperforms Complexpander for 26 of the
31 groups, while the converse never occurs (Additional data
file 3 and supplemental Figure S4 in [15]).
There are several reasons for the superior performance of
bioPIXIE. A major factor in its improvement is the robust
integration of a wide variety of genomic data. Both Asthana et
al[13] and Bader [14] focused their integration methodology
on physical interactions data (two-hybrid and affinity precip-
itation data). Our goal is to predict process-specific networks
rather than only complexes, which requires a more general
integration method applicable beyond physical interactions.
These diverse data types have varying degrees of information
across different complexes and processes, as evident from the
three KEGG pathways illustrated in Figure 1 and a broader

Page 4

R114.4 Genome Biology 2005, Volume 6, Issue 13, Article R114 Myers et al.
http://genomebiology.com/2005/6/13/R114
Genome Biology 2005, 6:R114
Figure 1 (see legend on next page)
00.10.20.30.40.50.6 0.70.80.91
0
0.2
0.4
0.6
0.8
1
Recall
( TP / [TP + FN] )
Precision
( TP/ [TP +FP] )
Integrated evidence
Physical association evidence
Genetic association evidence
Microarray correlation evidence
Performance of individual evidence types
00.1 0.20.3 0.40.50.6 0.70.80.91
0
0.2
0.4
0.6
0.8
1
Recall
( TP/[TP + FN] )
Precision
( TP/[TP + FP] )
bioPIXIE recovery
Binary recovery
Countingbased recovery
Comparison with naive methods
00.10.20.3 0.40.5 0.60.70.80.91
0
0.2
0.4
0.6
0.8
1
Recall
( TP / [TP + FN] )
Precision
( TP / [TP + FP] )
bioPIXIE recovery
SEEDY recovery
Complexpander recovery
Comparison with existing methods
0
5
10
15
20
25
Number of proteins recovered
0 50 100150200250300
0
0.2
0.4
0.6
0.8
1
Total graph size
Fraction of pathway recovered
Integrated evidence
Physical association evidence
Genetic association evidence
Microarray correlation evidence
0
10
20
30
40
50
60
70
80
Number of proteins recovered
0 50 100150200250300
0
0.2
0.4
0.6
0.8
1
Total graph size
Fraction of pathway recovered
Integrated evidence
Physical association evidence
Genetic association evidence
Microarray correlation evidence
0
10
20
30
40
50
60
70
80
90
100
110
120
130
140
Number of proteins recovered
050 100150200250300
0
0.2
0.4
0.6
0.8
1
Total graph size
Fraction of pathway recovered
Integrated evidence
Physical association evidence
Genetic association evidence
Microarray correlation evidence
Cell cycle (KEGG sce04110)
ATP synthesis (KEGG sce00193)
Ribosome (KEGG sce03010)
(a)
(b)
(c)
(d)
(e)
(f)

Page 5

http://genomebiology.com/2005/6/13/R114
Genome Biology 2005, Volume 6, Issue 13, Article R114 Myers et al. R114.5
comment
reviews
reports
refereed research
deposited research
interactions
information
Genome Biology 2005, 6:R114
study of bioPIXIE's performance on subsets of evidence (see
Additional data file 3). Our Bayesian integration can robustly
incorporate these data, which allows us to harness the infor-
mation from heterogeneous data types without sacrificing
specificity.
The search algorithm applied to the resulting integrated
probabilistic network is also a factor in bioPIXIE's improve-
ment over existing approaches. Our algorithm incorporates
information about both direct and indirect links between can-
didate proteins and the query set in a way that favors tightly
connected groups. SEEDY returns the weight of the maxi-
mum confidence link between a candidate protein and any
member of the query set, which only takes into account direct
connections and uses little information about the topology of
the network. Furthermore, the maximum is susceptible to
noise in both the query set and weights between pairs of pro-
teins. A single erroneous high-confidence link can bring a
candidate protein into the result set. The other algorithm
included for comparison, Complexpander, samples several
random binary networks whose edges are present with prob-
ability corresponding to the confidence in that interaction.
Proteins are ranked by the fraction of random networks in
which there exists a path, up to a maximum length (default of
four), from each protein to the query set. Although this algo-
rithm uses more information than SEEDY, both in terms of
topology and indirect links, we found its performance to scale
poorly with increased density of the weighted interaction net-
work. Specifically, as more genomic data are included in the
integration, the probabilistic integrated network becomes
more populated, resulting in many more possible (probability
>0) paths between any one protein and a particular query set.
There are so many paths that the fraction of random binary
networks with paths to the query set is no longer a discrimi-
native measure, which results in more false positives.
Although such a method might be appropriate for sparse
data, it does not appear to work well when larger datasets are
applied to the problem of query-based complex or pathway
recovery.
Another factor in the performance of our method is its robust-
ness to the quality and size of the query set. For each of the 31
groups of proteins described earlier, we evaluated the recov-
ery performance for 20 query proteins, of which between 1
and 19 were randomly chosen from the entire proteome and
the rest were chosen from the appropriate process or com-
plex. All 31 groups could tolerate 25% query set noise with less
than a 10% reduction in the average AUC; 27 of those could
tolerate 50% query set noise, and 14 of those could tolerate up
to 75% random proteins in the query set (see supplemental
Figure S5 in [15]). Thus, our method is robust to imperfect
query sets. We also evaluated the recovery performance over
a range of query set sizes from 4 to 60 proteins to determine
whether there was a noticeable decline in performance for
very small query sets. We found that, in general, the quality of
the network recovered from a pure query set of 4 to 5 proteins
is comparable to the result of a much larger query (40 to 50
proteins) on the same process, suggesting that relatively few
proteins are required to obtain a signal (supplemental Figure
S6 in [15]). For instance, with only a 4-protein query set,
bioPIXIE's maximum AUC score was within 10% of the max-
imum AUC score obtained on up to 60-protein query sets for
22 of the 31 processes (see supplemental Figure S6 in [15] for
supporting plot).
The query-driven nature of the search algorithm is a key fac-
tor in the accuracy of our method. The relationships between
query proteins selected by the user affect which neighboring
proteins are added to the final network. Thus, the network
resulting from a query is not simply a sub-section of the com-
plete integrated protein-protein interaction graph rooted at
the query proteins; rather, it is probabilistically biased by the
network search algorithm toward the specific biological con-
text represented in the query set. Figure 2 illustrates this
effect for the query protein Rad23. Rad23 is known to form a
complex with Rad4 (NEF2) and participate in nucleotide
excision repair [16]. Recent work has also suggested that
Rad23 facilitates DNA repair by inhibiting the degradation of
specific substrates in response to DNA damage [17,18].
Depending on which partners are included in a query with
Rad23, the network recovered by our system can focus on
Rad23's involvement in nucleotide excision repair or in ubiq-
uitin-dependent protein catabolism. For instance, when the
query includes DNA repair proteins Rad4, Rad3, and Rad24
bioPIXIE network recovery evaluation
Figure 1 (see previous page)
bioPIXIE network recovery evaluation. (a-c) Typical network recovery performance for three KEGG pathways. For all pathways, ten proteins from the
pathway were randomly picked as a query set. The results of 100 independent query set samplings are shown. The fraction of the total known process
components recovered is plotted versus the size of the graph grown from the query set. (d-f) An average over 31 KEGG pathways, GO biological
processes, and MIPS complexes. Performance is measured and reported as the trade-off between precision (the proportion of correct pathway
components returned to the total size of the returned network) and recall (the proportion of correct pathway components returned to the number of
total non-query pathway proteins). Precision and recall are derived from true positives (TP), false positives (FP), and false negatives (FN) as noted in the
axis labels. (d) The improvement gained by using our network prediction algorithm on a Bayesian integration of genomic evidence compared to separate
evidence types. bioPIXIE shows considerable improvement in both the number of known member proteins recovered and the precision of predicted
members for the integrated evidence over any individual evidence type. (e) The improved network recovery offered by the bioPIXIE algorithm versus
more naïve approaches to integration and graph search. Specifically, we plot the performance of bioPIXIE on integrated data against a naïve binary
approach for which information from all evidence types is used but only as a binary 'yes' or 'no' relationship, and a more sophisticated approach where
overlapping evidence receives higher weights and connected proteins are recovered in order of confidence. (f) Comparison of the performance of
bioPIXIE to two existing methods for query-based protein complex recovery [13,14].