MicroRNA Prioritization based on Target Profile Similarities
P´eter Marx
, Bence Bolg´ar
, Andr´as G´ezsi
, Attila Guly´as-Kov´acs
and P´eter Antal
Department of Measurement and Information Systems,
Budapest University of Technology and Economics, Budapest, Hungary
Department of Genetics, Cell- and Immunobiology, Semmelweis University, Budapest, Hungary
MTA-SE Neuropsychopharmacology and Neurochemistry Research Group, Budapest, Hungary
microRNA, microRNA Target, Kernel Methods, Multiple Kernel Learning, Gene Prioritization.
microRNAs form a complex regulatory network with thousands of target genes. This network is known to
suffer specific, but largely elusive, genetic perturbations in various types of disease. Accurate prioritization
of microRNAs for each disease type would elucidate those perturbations and so facilitate therapeutic and
diagnostic design. The multiple target profiles of microRNAs stemming from various experimental and in
silico methods allow the definition of wide range of similarities over microRNAs, but the combined use of
these of heterogeneous similarities was not utilized in the gene prioritization approach. Using microRNAs
as bases, prioritization with a disease-specific query set of microRNAs is straightforward once a microRNA-
microRNA similarity matrices have been derived. Here we demonstrate the application of a one-class version
of the multiple kernel learning framework in order to fuse heterogeneous characteristics of microRNAs. We
evaluate the method with breast cancer-specific queries, illustrate its technological aspects, and validate our
results not only by standard leave-one-out cross validation, but also with a prospective evaluation.
The growing availability of omic measurements and
multiple characterizations of entities like genes or
proteins led to the emergence of data and knowl-
edge fusion as a central challenge in various fields.
From the 90’s, the similarity based virtual screen-
ing became a popular method in chemoinformatics,
relying mostly on rank fusion and heuristic fusion
of data sources (Johnson and Maggiori, 1990; Ginn
et al., 1997; Ginn et al., 2000; Eckert and Bajorath,
2007). The aim of these methods were to make pre-
dictions based on different heterogeneousinformation
sources in various research fields. From the 2000s,
gene prioritization emerged as a separate task with
the fusion of wide range of information about genes
and gene products, such as sequence similarities,
transcriptional regulation similarities or expression
level similarities (Freudenberg and Propping, 2002;
Aerts et al., 2006; Kohler et al., 2008; Moreau and
Tranchevent, 2012). A theoretically sound foundation
for omic data fusion using similarities was proposed
in 2004 by Lanckriet et al., which utilized the kernel
methods for large-scale similarity integration (Lanck-
riet et al., 2004). Currently, wide range of methods
were reported for data and knowledge fusion, such
as the aforementioned kernel-based methods and the
more traditional text mining and network based fusion
methods (De Bie et al., 2007; Liekens et al., 2011; Lee
et al., 2011). Although there are many open questions,
such as the management of systematically incomplete
similarities, uncertainty in similarities, biological rel-
evance of ranks and proper evaluation, the kernel-
based fusion methods allows a universal framework
for any set of entities and it achieved an impressive
performance in multiple settings and domains, com-
pared to for example rank fusion (Bornigen et al.,
2012; Arany et al., 2013). Despite the universality
of the kernel-based fusion and prioritization frame-
work, the basis for the applications so far were nearly
exclusively gene centric, with some exceptions such
as its application for protein-ligand interaction pre-
diction (Iacucci et al., 2012) and for drug reposition-
ing (Arany et al., 2013; Bolg´ar et al., 2013). In this
paper we report the application of the kernel-based
fusion and prioritization framework over miRNAs, as
a step towards the extension of this methodology to
integrate information over heterogeneous sets of enti-
MiRNAs are short approximately 22 base pairs
Marx P., Bolgár B., Gézsi A., Gulyás-Kovács A. and Antal P..
MicroRNA Prioritization based on Target Profile Similarities.
DOI: 10.5220/0004925502780285
In Proceedings of the International Conference on Bioinformatics Models, Methods and Algorithms (BIOINFORMATICS-2014), pages 278-285
ISBN: 978-989-758-012-3
2014 SCITEPRESS (Science and Technology Publications, Lda.)
long non-coding RNAs and have an important role in
cis-regulation. Generally miRNAs recognize their tar-
get by the 2-8 bps long seed region at the 5’ end of the
miRNA. They bind to the 3’ end of the target mRNA
by Watson-Crick base pairing and can repress the
mRNA translation in multiple ways such as mRNA
degradation, mRNA de-deadenylation, inhibition of
initiation and premature ribosome drop off. MiRNAs
can also have binding sites in the coding region or at
the 5’ end of the targetmoreoverone miRNA can have
multiple targets and one target can bind to multiple
miRNAs. Although there is an exponentially increas-
ing interest towards miRNAs (more than 2000 miR-
NAs are known in human cells), plenty of their target
genes are only connected to them based on in silico
prediction. There is a wide variety of in silico meth-
ods for target exploration, prediction and validation,
but these methods are somewhat complementary and
suffer from different biases. To avoid these errors we
used only experimentally supported data in this paper.
For additional information on miRNAs and gene reg-
ulation see (Nilsen, 2007) and (Chen and Rajewsky,
Many studies suggested an important role for
miRNAs in the development of cancer based on the
observation that their expression levels changed in tu-
mor tissues which led to altered mRNA quantity in-
side the cell, and similarly both inherited and somatic
genetic variations related to miRNAs were reported as
associated with diseases. In the paper we illustrate the
new miRNA-based prioritization method with gen-
eral, disease aspecific kernels and with breast cancer
queries. Breast cancer is one of the leading cancer
types among the female population, and both miRNA
related genetic variations and expression levels were
reported to be associated with various phenotypic fea-
tures of breast cancer classes.
Our goal is to examine the efficiency of miRNA
prioritization based on the already available data
which could extend the current gene centric prioriti-
zation. Predicting new miRNA-disease connections
can lead to a more detailed description of the regula-
tion of the disease in focus. Since the miRNA data is
coming from different experiments and including in
silico target predictions would increase heterogeneity
we use a well known information fusion method as
well and compare it to the other one kernel prioriti-
zation algorithm. For evaluation of the performance
of the prioritization we apply the leave-one-out cross-
validation and because of its potential positive bias
through knowledge contamination, see e.g. (Kohler
et al., 2008; Bornigen et al., 2012), we performed a
prospective evaluation using a recent summary from
Jacobsen et al. (Jacobsen et al., 2013).
The data fusion based prioritization methods through
integration of similarities and rankings from multi-
ple information source has a long and somewhat re-
dundant history, with its roots in virtual screening
in chemoinformatics (Johnson and Maggiori, 1990;
Ginn et al., 1997; Ginn et al., 2000; Eckert and
Bajorath, 2007), kernel methods in chemoinformat-
ics (Burbidge et al., 2001; Warmuth et al., 2003), in
the one-class classification and prioritization (Moya
and Hush, 1996; Schlkopf et al., 2001), in multi-
ple kernel learning (Lanckriet et al., 2004; Rako-
tomamonjy et al., 2008) and in gene prioritization
methods (Freudenberg and Propping, 2002; Aerts
et al., 2006; Kohler et al., 2008; De Bie et al., 2007;
Liekens et al., 2011; Lee et al., 2011; Moreau and
Tranchevent, 2012).
Compared to the growing admittance of the cen-
tral role of miRNAs, the use of this information
source is surprisingly rare. For example, transcrip-
tion regulation based gene-gene similarities (kernels)
are dominantly constructed base on the commonality
of the transcription factor binding sites of the gene
pairs, and not by the commonality of miRNAs (see
e.g. (Aerts et al., 2006)). Two notable exceptions re-
ported the prioritization of miRNAs using a network
approach (Xu et al., 2011) and using a support vector
machine classifier (Li et al., 2012). The work reported
in this paper continues this line as it uses miRNAs for
the bases of prioritization, instead of the prevailing
gene centric approach. However, this work adopts a
more principled approach for fusion by adopting the
multiple kernel learning framework in the one-class
classification (/prioritization) settings.
The number of databases containing miRNA-target
pairs is growing just like the methods which utilize
machine learning algorithms to find further potential
targets for a miRNA (Ficarra et al., 2012). However
applying such methods largely increase the number
of miRNA-target pairs and extend the miRNA set us-
ing predicted connections, it leads to higher degree of
uncertainty at the results. Therefore we used only ex-
perimentally supported targets for every miRNA. Us-
ing the predictions based results requires an extensive
evaluation of the methods which will be aimed in a
future study.
3.1 Computing miRNA Kernel
We used miRTarBase (Hsu et al., 2011) (downloaded
11/14/2013 Release: 4.5) to collect experimentally
validated miRNA-target pairs. We filtered out the
non-human target genes from the data resulting in a
dataset of 596 unique miRNAs and their experimen-
tally supported targets. Similarity between two miR-
NAs is defined by the number of common target genes
in a given context, e.g. using a given miRNA target
validation method. The primary diagonal of the ma-
trix X contains the degree of a miRNA which equals
with the number of targets in the present study. To
extend our definition of miRNA similarity and to en-
sure positive definiteness we squared the matrix and
normalized it with a diagonal matrix D containing the
reciprocal of the square root of the diagonal elements
of X.
K = DX
This way two miRNAs i, j will be more similar (x
will be higher) if they share more targets or they have
common similar miRNAs. miRTarBase contains in-
formation on the experimental methods used to vali-
date the miRNA-target pairs. These methods can be
classified into two groups based on the strength of the
evidence they provide. E.g. microarrays give indi-
rect proof as the prediction is based on the differential
expression miRNA-gene pairs therefore it is a ’weak’
validation. On the contrary attaching reporter genes
to the target gene proves directly the miRNA effects
on the gene (Ficarra et al., 2012). We used the classi-
fication of miRTarBase (functional, functional-weak,
nonfunctional, nonfunctional-weak) and built differ-
ent kernels for each of them with the above described
3.2 Multiple Kernel Learning
The most important benefit of Multiple Kernel Learn-
ing is that it provides a way to perform both data fu-
sion and prioritization in a joint manner, by solving
one single optimization problem. This is achieved
by incorporating kernel “weights” into the objec-
tive function, which essentially amounts to determin-
ing an optimal weighting of (possibly very heteroge-
neous) information sources by exploiting the infor-
mation content of the query itself. Further favorable
properties include computational efficiency, flexibil-
ity (depending solely on pairwise similarities), and
often very good empirical performance. Multiple ker-
nel methods have already been successfully applied in
the context of genomic data fusion (Lanckriet et al.,
2004), however, first formulations did not support
gene prioritization. The first tool which performed
fusion and prioritization in a coupled way involved a
modification of the multiple kernel one-class SVM al-
gorithm (Sch¨olkopf et al., 2001) to perform the rank-
ing of the entities based on their projections to the
normal of the hyperplane (De Bie et al., 2007). In
order to attain better computational performance, we
modified the formulation of Vishwanathan et al. (Sun
et al., 2010):
k f
) ρ ξ
ξ 0, d
d 0, i = 1, 2, . . . , l,
where f
f parameterizes the hyperplane, d
is the
weight of the kth kernel, ρ denotes the margin, ν con-
trols the model complexity, ξ
ξ are the slack variables
and the last term stands for the Lp-norm regulariza-
tion of the kernel weights. Note that this is just the
primal of the one-class SVM augmented by the ker-
nel weights.
In view of earlier findings (Yu et al., 2010), we
utilized the squared L2-norm as weight regularizer.
In this setting, the dual reads
s.t. 0 α
1, 1
α = νl,
where α
α are the dual variables and the optimal ker-
nel weights can be recovered from the solution. Note
that this dual is differentiable and convex, which is
very easy to optimize. The score of a sample can be
computed as
x) =
, x
i.e. here we determine the distance to the hyperplane
instead of the side on which the sample x
x lies; samples
are then ranked on the basis of their respective scores.
We used multiple kernels and training sets (queries)
to cover the different subtypes of breast cancer clas-
sified by Samantarrai et al. (Samantarrai et al., 2013)
as follows: Ductal carcinoma in situ, Lobular carci-
noma in situ, Invasive ductal- or invasive lobular car-
cinoma. Our first approach was to collect a subset
of miRNAs from one subtype of breast cancer and
1 5 20 100 500
0.0 0.2 0.4 0.6 0.8
Figure 1: The scores of the miRNAs for the MKL kernel
and the third query (axis X is log scaled).
one type of differential expression (up- or downreg-
ulated). We will refer to these subset of miRNAs as
a class. This way the prioritization should recover
the remaining set of miRNAs of the same class, and
it should give a higher score to other miRNAs con-
nected with breast cancer. The other approach was
to choose miRNAs which are present in more sub-
classes to catch the different properties of each sub-
type and prioritize with this breast cancer “hyper-
class”. In both cases we prioritize with the MKL
algorithm and also the kernel built from the full hu-
man miRTarBase. The first query (Query1) contained
the hsa-miR-182, hsa-miR-183, hsa-miR-200c, hsa-
miR-21. The second query (Query2) contained miR-
NAs of the subtype ductal carcinoma in situ downreg-
ulated miRNAs (hsa-miR-125b, hsa-miR-127, hsa-
miR-210, hsa-miR-7). We included those miRNAs in
the last query (Query3) which were present in more
classes to span the set of miRNAs which are con-
nected to the disease. These are hsa-let-7d, hsa-miR-
182, hsa-miR-183, hsa-miR-21, hsa-miR-210, hsa-
miR-221. When it was not specified and for one pre-
miRNA more mature miRNA were available we used
both the 3p and 5p miRNA. We analyzed other queries
also which gave similar results and for this reason
those are not included in the manuscript.
Query1 with the full kernel ranked the other miR-
NAs in the same class hsa-miR-361-5p, hsa-miR-
374a, hsa-miR-93 113th, 52th, 12th respectively and
resulted an average 126.46 rank for the all miRNAs
whereas the MKL kernel ranked them 29th, 84th, 13th
with an average of 121.96. Both methods ranked
miRNAs from other classes in the first 10% of the
Table 1: Leave-one-out cross validation results for the duc-
tal carcinoma in situ breast cancer type upregulated miR-
NAs. The numbers indicate the rank of the left out miRNA.
microRNA Full MKL
hsa-miR-21 169 113
hsa-miR-200c 257 57
hsa-miR-182 41 31
hsa-miR-183 60 54
hsa-miR-361-5p 101 28
hsa-miR-374a 63 119
hsa-miR-93 16 14
Average 101 59.43
data. The rank of the miRNA is in brackets. Full:
hsa-miR-221 (8), hsa-miR-96 (44); MKL: hsa-miR-
221 (15), hsa-miR-96 (42), hsa-miR-18 (44). Query2
ranked the miRNAs from the same class (first num-
ber belongs to the full kernel and the second to the
MKL kernel set) hsa-let-7d (222,123), hsa-miR-221
(88,60), hsa-miR-320 (23,66) and others full: hsa-
miR-182 (51), hsa-miR-361-5p (43), hsa-miR-9 (50),
hsa-miR-93 (20) and MKL: hsa-miR-182 (25), hsa-
miR-361-5p (33), hsa-miR-9 (41), hsa-miR-93 (38),
hsa-miR-10b (40) and hsa-miR-21 (57). The average
ranks of miRNAs from all subtypes are 120.83 for the
full kernel and 112.92 for the MKL set. The results
for the last query is similar to the first two, generally
using MKL kernel set givesa lower average rank. Fig-
ure 1 shows the scores of the last query for the MKL
kernel set on a log scale. The cut-off point shows a
break between the query elements and the other miR-
NAs. The MKL method computes a weight for every
kernel. In our case these weights were almost uniform
and not varied between the different queries. Gener-
ally, the weak functional set had 0.4 weight and the
other kernels 0.2.
For validation we used not only the same set of
miRNA, but also the results from Jacobsen et al.
(Jacobsen et al., 2013) for a prospective evaluation.
They computed the correlation R
between miRNAs
expression levels and DNA copy number variation.
Based on their results the variation of the expression
levels in case of cancer related miRNAs can be ex-
plained by DNA copy number variation. We selected
the miRNAs which has a higher than 0.1 correlation
and which are known to play a role in breast can-
cer. This routine makes it possible to validate our
results against a newer breast cancer dataset (see Ta-
ble 2). Furthermore, leave-one-out cross validation
(LOOCV) was performed on the Ductal carcinoma in
situ upregulated miRNA set. The average rank of the
left out miRNA was 101 with the full kernel and 59.43
with the MKL algorithm (see Table 1).
Figure 2: The graph of the first 50 ranked miRNAs for the full kernel and the first query. Only edges with higher than 0.4
similarity scores are shown. The color of a node is defined by its score, where blue indicates lower score.
Table 2: The miRNAs ranked below 50 from the Jacobsen
dataset. *The miRNA was part of the query.
Query1 Query3
Full MKL Full MKL
hsa-miR-106b 11 51 12 79
hsa-miR-141 51 27 375 386
hsa-miR-15b 26 11 47 30
hsa-miR-17 41 20 85 41
hsa-miR-19b 13 24 34 36
hsa-miR-200c 6* 5* 280 95
hsa-miR-20a 43 28 92 88
hsa-miR-21 2* 2* 14* 3*
hsa-miR-92a 16 31 3 18
hsa-miR-140 61 22 121 80
hsa-miR-182 3* 6* 16* 2*
hsa-miR-183 1* 1* 15* 1*
hsa-miR-186 76 18 68 34
hsa-miR-25 10 25 9 37
hsa-miR-320a 107 87 32 43
Despite the popularity of gene prioritization, there are
only a few studies in which researchers used miRNA
prioritization for finding novel disease related miR-
NAs. To our knowledge, those are only validated by
LOOCV. The quantitative evaluation of the perfor-
mance of our method for the queries (Query 1,2,3)
shows that it ranks most of the miRNAs related to
breast cancer according to the review from Samantar-
rai et al. (Samantarrai et al., 2013) in the first 20%
of the miRNA set used in our study. Bornigen et
al. (Bornigen et al., 2012) used the top 5 %, 10 %,
30 % to validate the performance of the prioritiza-
tion methods for real examples. Taking the number
of miRNAs into account this rate (top 20% includes
almost all miRNAs related to breast cancer) can be
enough for experimental validation with a real situa-
tion with limited budget also.
On Figure 2 miRNAs are connected based on the
similarity of miRNA-miRNA pairs. Nodes even with
lower rank has plenty of neighbors, on the other hand
nodes with higher scores, typically query members,
share less properties with the other miRNAs in the
graph (i.e. connected with edge above the threshold
0.4). The manual evaluation of the graph also sug-
gested that the prioritization gives higher score not
only for those miRNAs which are in the same miRNA
cluster as the query miRNAs. Generally the nodes
with more connections belongs to a miRNA family
represented by more members in the first 50 ranked
entities, which verifies our similarity definition.
The prospective validation on the results of Ja-
cobsen et al. ranked the miRNAs in Table 2 on
the first 10% of the whole set. This means that
the applied method was able to identify new can-
didates which were experimentally validated later
without retrospective bias. The role of these miR-
NAs in cancer development is supported by exper-
imental evidence. For example miR-17 and miR-
200c are oncosuppressors and miR-92 is an oncomiR
(Samantarrai et al., 2013). Furthermore (Kim et al.,
2011) found inverse correlation between hsa-miR-17,
hsa-miR-92, hsa-miR-106b, hsa-miR-20a, hsa-miR-
19 and ZBTB4 gene which has a significant corre-
lation to relapse-free survival. The almost uniform
kernel weights most likely caused by the similar in-
put data as for plenty of miRNA-target pairs are sup-
ported with more than one experimental evidence.
The higher weight of weak functional kernel is pos-
sibly the consequence of the higher number of con-
nections available. Our results show that the MKL
method which includes both prioritization and fusion
outperformed the single kernel prioritization in all of
the above cases. With even more heterogeneous in-
formation source the application of the MKL method
can be even more advantageous.
The high heterogeneity of experimental and in sil-
ico methods for miRNA target exploration, prediction
and validation result in highly heterogeneous miRNA
profile and respectively, miRNA similarities. We ap-
plied the multiple kernel learning framework for the
fusion of these similarities in breast cancer. Results
indicate that this methodology can be a promising ap-
proach to find new miRNAs related to cancer.
Furthermore, this approach could be used to re-
fine existing miRNA families and to define more de-
tailed gene-gene similarities . Our results also indi-
cate that miRNA prioritization can support study de-
sign and interpretation of miRNA discovery, and it
also indicates the universal applicability of kernel-
based fusion methods over different sets of entities,
such as miRNAs, beside the current gene centric ap-
proach. With the growing number of discovered miR-
NAs and the deeper understanding of their role in the
post-transcriptional regulation miRNA prioritization
can lead to a more detailed description of miRNA-
gene regulatory networks.
However, the approach described in this paper
seems to be successfully identify even novel miR-
NAs for breast cancer still there are many question
to be answered. MiRNA networks or miRNA simi-
larity can be defined in other ways such as comparing
the sequence or based on single nucleotide polymor-
phisms (SNP) or by using conservation scores. After
a thorough investigation of the target prediction algo-
rithms we can include predicted miRNA-target pairs
in the analysis to increase the number of miRNAs in
the study. Last with the growing number of avail-
able data it will be possible soon to build more tissue
specific miRNA networks. This way we will have a
heterogeneous data source and we can make a good
use of the MKL method as it can be used to com-
pare the importance of the different similarity scores
or data sources based on the learned weights of the
kernels. Future work will concentrate on answering
these questions.
Aerts, S., Lambrechts, D., Maity, S., Van Loo, P., Coessens,
B., De Smet, F., Tranchevent, L., De Moor, B., Mary-
nen, P., Hassan, B., Carmeliet, P., and Moreau, Y.
(2006). Gene prioritization through genomic data fu-
sion. Nature Biotechnology, 24(5):537–544.
Arany, A., Bolg´ar, B., Balogh, B., Antal, P., and M´atyus,
P. (2013). Multi-aspect candidates for repositioning:
Data fusion methods using heterogeneous information
sources. Current Medicinal Chemistry, 20(1):95–107.
Bolg´ar, B., Arany, A., Temesi, G., Balogh, B., Antal, P., and
M´atyus, P. (2013). Drug repositioning for treatment
of movement disorders: from serendipity to rational
discovery strategies. CURRENT TOPICS IN MEDIC-
INAL CHEMISTRY, 13(18):2337–63.
Bornigen, D., Tranchevent, L., Bonachela-Capdevila, F.,
Devriendt, K., De Moor, B., De Causmaecker, P., and
Moreau, Y. (2012). An unbiased evaluation of gene
prioritization tools. Bioinformatics, 28(23):3081–
Burbidge, R., Trotter, M., Buxton, B., and Holden, S.
(2001). Drug design by machine learning: sup-
port vector machines for pharmaceutical data analysis.
Comput. Chem. (Oxford, U. K.), 26(1):5–14.
Chen, K. and Rajewsky, N. (2007). The evolution of gene
regulation by transcription factors and micrornas. Nat
Rev Genet, 8(2):93–103.
De Bie, T., Tranchevent, L.-C., Van Oeffelen, L. M., and
Moreau, Y. (2007). Kernel-based data fusion for gene
prioritization. Bioinformatics, 23(13):i125–i132.
Eckert, H. and Bajorath, J. (2007). Molecular similarity
analysis in virtual screening: foundations, limitations
and novel approaches. Drug Discovery Today, 12(5-
Ficarra, E. et al. (2012). One decade of development and
evolution of microrna target prediction algorithms.
Genomics, proteomics & bioinformatics.
Freudenberg, J. and Propping, P. (2002). A similarity-
based method for genome-wide prediction of disease-
relevant human genes. Bioinformatics, 18:S110–
Ginn, C., Turner, D., Willett, P., Ferguson, A., and Her-
itage, T. (1997). Similarity searching in files of three-
dimensional chemical structures: Evaluation of the
eva descriptor and combination of rankings using data
fusion. J. Chem. Inf. Model., 37(1):23–37.
Ginn, C., Willett, P., and Bradshaw, J. (2000). Combina-
tion of molecular similarity measures using data fu-
sion. Perspect. Drug Discovery Des., 20(1):1–16.
Hsu, S.-D., Lin, F.-M., Wu, W.-Y., Liang, C., Huang, W.-
C., Chan, W.-L., Tsai, W.-T., Chen, G.-Z., Lee, C.-J.,
Chiu, C.-M., et al. (2011). mirtarbase: a database cu-
rates experimentally validated microrna–target inter-
actions. Nucleic acids research, 39(suppl 1):D163–
Iacucci, E., Tranchevent, L., Popovic, D., Pavlopoulos, G.,
De Moor, B., Schneider, R., and Moreau, Y. (2012).
Reliance: a machine learning and literature-based pri-
oritization of receptor-ligand pairings. Bioinformat-
ics, 28(18):I569–I574.
Jacobsen, A., Silber, J., Harinath, G., Huse, J. T., Schultz,
N., and Sander, C. (2013). Analysis of microrna-target
interactions across diverse cancer types. Nature struc-
tural & molecular biology.
Johnson, M. and Maggiori, G. (1990). Concepts and appli-
cations of molecular similarity. Wiley.
Kim, K., Chadalapaka, G., Lee, S., Yamada, D., Sastre-
Garau, X., Defossez, P.-A., Park, Y.-Y., Lee, J.-S., and
Safe, S. (2011). Identification of oncogenic microrna-
17-92/zbtb4/specificity protein axis in breast cancer.
Oncogene, 31(8):1034–1044.
Kohler, S., Bauer, S., Horn, D., and Robinson, P. (2008).
Walking the interactome for prioritization of candi-
date disease genes. American Journal of Human Ge-
netics, 82(4):949–958.
Lanckriet, G., De Bie, T., Cristianini, N., Jordan, M.,
and Noble, W. (2004). A statistical framework for
genomic data fusion. Bioinformatics, 20(16):2626
Lee, I., Blom, U., Wang, P., Shim, J., and Marcotte,
E. (2011). Prioritizing candidate disease genes by
network-based boosting of genome-wide association
data. Genome Research, 21(7):1109–1121.
Li, X., Wang, Q., Zheng, Y., Lv, S., Ning, S., Sun, J.,
Huang, T., Zheng, Q., Ren, H., Xu, J., Wang, X.,
and Li, Y. (2012). Prioritizing human cancer mi-
crornas based on genes? functional consistency be-
tween microrna and cancer. Nucleic Acids Research,
Liekens, A., De Knijf, J., Daelemans, W., Goethals, B.,
De Rijk, P., and Del-Favero, J. (2011). Biograph: un-
supervised biomedical knowledge discovery via auto-
mated hypothesis generation. Genome Biology, 12(6).
Moreau, Y. and Tranchevent, L. (2012). Computational
tools for prioritizing candidate genes: boosting dis-
ease gene discovery. Nature Reviews Genetics,
Moya, M. and Hush, D. (1996). Network constraints and
multi-objective optimization for one-class classifica-
tion. Neural Networks, 9(3):463–474.
Nilsen, T. W. (2007). Mechanisms of microrna-mediated
gene regulation in animal cells. TRENDS in Genetics,
Rakotomamonjy, A., Bach, F., Canu, S., and Grandvalet, Y.
(2008). Simplemkl. J. Mach. Learn. Res., 9:2491–
Samantarrai, D., Dash, S., Chhetri, B., and Mallick, B.
(2013). Genomic and epigenomic cross-talks in
the regulatory landscape of mirnas in breast cancer.
Molecular Cancer Research, 11(4):315–328.
Schlkopf, B., Platt, J., Shawe-Taylor, J., Smola, A., and
Williamson, R. (2001). Estimating the support of
a high-dimensional distribution. Neural Comput.,
Sch¨olkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J.,
and Williamson, R. C. (2001). Estimating the support
of a high-dimensional distribution. Neural computa-
tion, 13(7):1443–1471.
Sun, Z., Ampornpunt, N., Varma, M., and Viswanathan, S.
(2010). Multiple kernel learning and the smo algo-
rithm. In Advances in neural information processing
systems, pages 2361–2369.
Warmuth, M., Liao, J., Ratsch, G., Mathieson, M., Putta,
S., and Lemmen, C. (2003). Active learning with sup-
port vector machines in the drug discovery process. J.
Chem. Inf. Model., 43(2):667–673.
Xu, J., Li, C.-X., Lv, J.-Y., Li, Y.-S., Xiao, Y., Shao, T.-T.,
Huo, X., Li, X., Zou, Y., Han, Q.-L., Li, X., Wang,
L.-H., and Ren, H. (2011). Prioritizing candidate dis-
ease mirnas by topological features in the mirna tar-
get?dysregulated network: Case study of prostate can-
cer. Molecular Cancer Therapeutics, 10:1857–1866.
Yu, S., Falck, T., Daemen, A., Tranchevent, L.-C., Suykens,
J. A., De Moor, B., and Moreau, Y. (2010). L2-norm
multiple kernel learning and its application to biomed-
ical data fusion. BMC bioinformatics, 11(1):309.