PACCANAROLAB
Computational selection
of transcriptomics experiments improves Guilt-by-Association analyses
Prajwal Bhat1,2,3, Haixuan Yang2,3, Laszlo Bögre1,3,
Alessandra Devoto1,3* and Alberto Paccanaro2,3*
1School
of Biological Sciences, Royal Holloway University of London, Egham, United
Kingdom
2Department
of Computer Science, Royal Holloway University of London, Egham, United Kingdom
3Centre
for Systems and Synthetic Biology, Royal Holloway University of London, Egham,
United Kingdom
*Corresponding
authors
Microarray
is perhaps the most widely used gene expression measurement technology today.
Gene expression profiles obtained from microarray experiments provide an
insight into gene function and it is considered a powerful exploratory
technique. However, understanding gene function from the expression profiles is
a challenge. One of the widely used principles for elucidating the function of
genes is the Guilt by Association (GBA) principle where, genes are considered
to be involved in similar functions if the gene expression profiles are
correlated. GBA principle is used in a variety of functional analyses
techniques ranging from simple co-expression analysis, clustering techniques to
network based approaches.
Generally in gene functional
analyses, the GBA principle is applied to large collections of microarrays from
various biological backgrounds such as different treatments and tissues.
However, it is generally accepted that using large heterogeneous collections of
microarrays could sometimes be unreasonable due to the condition specific
nature of gene expression. This is evident when we observe the distribution of
correlation coefficients among genes which are involved in the same function
(See figure 1A below). If genes in the same functional category show poor
correlation, then based on correlation, identifying new genes which might be
involved in that function would be not possible.
Figure 1: Distribution of
correlation coefficients for genes in the GO category GO:009753 “Response to Jasmonic
Acid stimulus” calculated using (a) a large heterogeneous collection of
experiments and (b) a manually selected set of experiments, which were deemed
to be functionally relevant to Jasmonic Acid response based on literature
knowledge. To account for the different vector length in the two cases,
correlation coefficients were counted only when significant (p-value<0.05).
In our work, we have shown that
using only those experiments which are relevant to the functional category of
interest could greatly improve the correlation between the genes in the same
functional category (figure 1B). An obvious way to identify experiments that
may be relevant to the analysis is by simply reading the literature. However, with ever increasing size of
microarray compendia, identifying relevant experiments by literature review
alone is an impractical task. In our paper, we present an algorithm that
identifies functionally relevant experiments from a large collection of
microarrays.
Figure 3: Distribution
of correlation coefficient in GO BP terms show that the selected set of
experiments enrich higher correlations when compared to using all experiments.
The p-values from the t-test indicate that the distribution of
correlations for the selected set is significantly greater than for all
experiments.
We find that the experiments
selected by our algorithm lead to a significantly higher distribution of
correlation between the genes which belong to the same functional category when
compared to using all experiments in the compendia (See Figure 2). We also find
that a GBA based classifier will consistently outperform when experiments
selected by our algorithm are used for calculating the correlation compared to
using all experiments. Interestingly, the performance of the GBA based
classifier when selected experiments are used varies with specificity of the
functional annotation i.e. the performance is better with the more specific functional
categories (lower nodes of the GO Biological process tree) compared to the
broader functional categories (higher nodes in the GO tree).
The MATLAB code for the algorithm
and the data used in the paper can be downloaded from here