user login 

BITS2007 Meeting
BITS2007 Meeting



26-28 April 2007 Napoli, Italy

email support
Home > Contribution details
get PDF of this contribution get XML of this contribution get ICal of this contribution
 
 
 
A computationally efficient method for gene network identification
 
Motivation
An increasing amount of data from gene expression array experiments is
becoming available due to increasing interest toward gene regulatory
networks and the decreasing costs of the DNA chip technology. A number
of approaches have been proposed to ``reverse engineer`` such
data. All known inference methods are based on statistical techniques
and are characterized by a trade-off between selectivity (ability to
discriminate between true inferred gene interactions from false ones)
and sensitivity (percentage of true interactions
recovered). Currently, no methods exist that can recover more than a
small fraction of the regulatory interactions with high confidence;
high confidence is desirable because the in vitro verification of
putative interactions is expensive. A new promising approach is based
on combining a statistical inference method and an iterative algorithm
in which a solution is obtained through successive approximations. In
order to make this approach feasible, a computationally fast inference
method needs to be developed.



Methods
In this work we focus on a two-stage approach in which we first select
a subset of genes whose expression profiles appear to be correlated,
and then apply a method to estimate the coupling coefficients for the
selected genes based on their profiles. For the second step we chose a
fast and theoretically elegant statistical inference method based on
entropy maximization proposed by Lezon et al.~\cite{Lezon2006}. The
first step of gene selection is the most critical and less
theoretically characterized of the two; we focused most of our efforts
on this step. We empirically tested several different approaches and
compared their performance using a synthetic data set of 100 genes and
100 experiments.  The synthetic data set was the same used in the
review paper by Bansal et al.~\cite{Bansal2007}, therefore we were
able to directly compare our results with those obtained with
mainstream network inference tools.
 
The first selection gene rule we tested was a straightforward
implementation of the selection rule presented in the Lezon et
al. paper~\cite{Lezon2006}, based on the selection of genes whose
expression profiles have values well above the average value. We then
improved on the original approach by using selection criteria based on
the value of correlation coefficient between pairs of gene
expressions; more specifically, we tried using the values of the
coefficients themselves, and then the associated p-values. The rule
based on p-values turned out to be the one giving the best results.
We tried other approaches to selection of genes, inspired to methods
found in the literature on gene regulatory network inference. Briefly,
for one of these rules we computed the entropy of each gene expression
profile; the computation was performed using an histogram of the
expression values. In another rule we tried to apply the CLR algorithm
to the pre-filtering of genes, instead of post filtering of coupling
coefficients as originally proposed~\cite{Faith2007}. In the last two
rules we used the value of Mutual Information (MI) as the selection
criteria; MI is computed using the histogram method and the kernel
density estimation method respectively, as described in
\cite{Steuer2002}.  




Results
In this paper we propose a fast method for inferring genetic
interaction networks from gene expression data. By decomposing the
problem of inferring a network into two sub-steps, gene selection and
coupling coefficient determination, we were able to use simpler
computational methods than those currently proposed in the literature.
Based on an evaluation with synthetic data, our two-stage approach is
between 2 and 3 orders of magnitude faster than the best published
method and only a factor of 2.5 to 3 worse in terms of selectivity for
a comparable level of sensitivity.
 
Id: 132
Place: Napoli, Italy
Centro Congressi "Federico II"
Via Partenope 36
Napoli
Starting date:
-- not yet scheduled --   
Duration: 01h00'
Contribution type: Poster
Primary Authors: LAURIA, Mario (TIGEM, Napoli)
Co-Authors: DI BERNARDO, Diego (TIGEM, Napoli)
Presenters: LAURIA, Mario
 
Included in session: Poster Session
Included in track: Gene expression and system biology
 




bits2007_support@ceinge.unina.it | Last modified 08 July 2009 10:35 |




Powered by Indico 0.90.3