bbc 2015

BeNeLux Bioinformatics Conference – Antwerp, December 7-8 2015
Abstract ID: O6
Oral presentation
10th Benelux Bioinformatics Conference bbc 2015
O6. COMBINING TREE-BASED AND DYNAMICAL SYSTEMS
FOR THE INFERENCE OF GENE REGULATORY NETWORKS
Vân Anh Huynh-Thu 1* & Guido Sanguinetti 2,3 .
GIGA-R & Department of Electrical Engineering and Computer Science, University of Liège 1 ; School of Informatics,
University of Edinburgh 2 ; SynthSys – Systems and Synthetic Biology, University of Edinburgh 3 . * vahuynh@ulg.ac.be
INTRODUCTION
Reconstructing the topology of gene regulatory networks
(GRNs) from time series of gene expression data remains
an important open problem in computational systems
biology. Current approaches can be broadly divided into
model-based and model-free approaches, and face one of
two limitations: model-free methods are scalable but
suffer from a lack of interpretability, and cannot in general
be used for out of sample predictions. On the other hand,
model-based methods focus on identifying a dynamical
model of the system; these are clearly interpretable and
can be used for predictions, however they rely on strong
assumptions and are typically very demanding
computationally. Here, we aim to bridge the gap between
model-based and model-free methods by proposing a
hybrid approach to the GRN inference problem, called
Jump3 (Huynh-Thu & Sanguinetti, 2015). Our approach
combines formal dynamical modelling with the efficiency
of a nonparametric, tree-based method, allowing the
reconstruction of GRNs of hundreds of genes.
METHODS
Gene expression model. At the heart of the Jump3
framework, we use the on/off model of gene expression
(Ptashne & Gann, 2002), where the rate of transcription of
a gene can vary between two levels depending on the
activity state μ of the promoter of the gene. The expression
x of a gene is modelled through the following stochastic
differential equation:
dx i = (A i μ i (t) + b i – λ i x i )dt + σdω(t),
where subscript i refers to the i-th target gene. Here, the
promoter state μ i (t) is a binary variable (the promoter is
either active or inactive) that depends on the expression
levels of the transcription factors (TFs) that bind to the
promoter. A i , b i and λ i are kinetic parameters, and the term
σdω(t) represents a white noise-driving process with
variance σ 2 .
Network reconstruction with jump trees. Recovering
the regulatory links pointing to gene i amounts to finding
the genes whose expression is predictive of the promoter
state μ i . To achieve this goal, we propose a procedure that
learns, for each target gene i, an ensemble of decision trees
predicting the promoter state μ i at any time t from the
expression levels of the candidate regulators at the same
time t. However, standard tree-based methods cannot be
applied here since the output μ i (t) is a latent variable. We
therefore propose a new decision tree algorithm called
“jump tree”, which splits the observations by maximising
the marginal likelihood of the dynamical on/off model.
The learned tree-based model is then used to derive an
importance score for each candidate regulator, computed
as the sum of the likelihood gains that are obtained at all
the tree nodes where this regulator was selected to split the
observations. The importance of a candidate regulator j is
used as weight for the putative regulatory link of the
network that is directed from gene j to gene i.
RESULTS & DISCUSSION
We evaluated Jump3 on the networks of the DREAM4 In
Silico Network challenge (Prill et al., 2010). For each
network topology, two types of simulated expression data
were used: data simulated using the on/off model (toy
data) and the time series data that was provided in the
context of the DREAM4 challenge. We compared Jump3
to other GRN inference methods: two model-free methods,
which are time-lagged variants of GENIE3 (Huynh-Thu et
al., 2010) and CLR (Faith et al., 2007) respectively; two
model-based methods, namely Inferelator (Greenfield et
al., 2010) and TSNI (Bansal et al., 2006), and G1DBN
(Lèbre, 2009), a method based on dynamic Bayesian
networks. Areas Under the Precision-Recall curves
(AUPRs) obtained for size-100 networks are shown in
Table 1. Jump3 yields the highest AUPR in the case of the
toy data. As expected, its performance decreases when the
networks are inferred from the DREAM4 data, due to the
mismatch between the on/off model and the one used to
simulate the data. However, Jump3 still outperforms the
other methods.
Toy
DREAM4
Jump3 0.272 ± 0.060 0.187 ± 0.058
GENIE3-lag 0.114 ± 0.010 0.176 ± 0.056
CLR-lag 0.088 ± 0.008 0.169 ± 0.047
Inferelator 0.069 ± 0.006 0.144 ± 0.036
TSNI 0.020 ± 0.003 0.042 ± 0.010
G1DBN 0.104 ± 0.024 0.114 ± 0.043
TABLE 1. Comparison of network inference methods (mean AUPR and
standard deviation).
We also applied Jump3 to gene expression data from
murine bone marrow-derived macrophages treated with
interferon gamma (Blanc et al., 2011). Several of the hub
TFs in the predicted network have biologically relevant
annotations. They include interferon genes, one gene
associated with cytomegalovirus infection, and cancerassociated
genes, showing the potential of Jump3 for
biologically meaningful hypothesis generation.
REFERENCES
Bansal M et al. Bioinformatics 22, 815-822 (2006).
Blanc M et al. PLoS Biol 9, e1000598 (2011).
Faith JJ et al. PLoS Biol 5, e8 (2007).
Greenfield A. PLoS ONE 5, e13397 (2010).
Huynh-Thu VA & Sanguinetti G. Bioinformatics 31, 1614-1622 (2015).
Huynh-Thu VA et al. PLoS ONE 5, e12776 (2010).
Lèbre S. Stat Appl Genet Mol Biol 8, Article 9 (2009).
Prill RJ et al. PLoS ONE 5, e9202 (2010).
Ptashne M & Gann A. Genes and Signals. Cold Harbor Spring
Laboratory Press (2002).
26