Hello,
i have datasets derived from microarray experiments and I want to reconstruct gene coexpression networks.I know i have to compute the Pearson Correlation coefficient at first. Are there different methods to continue the analysis and reconstruct the networks? Can you recommend software?
Book, description, etc: http://www.genetics.ucla.edu/labs/horvath/CoexpressionNetwork/
R package: http://www.genetics.ucla.edu/labs/horvath/CoexpressionNetwork/Rpackages/WGCNA/
I wonder whether a new correlation method instead of that of Pearson's can be useful here:
There are several useful Cytoscape plugins. This tutorial might be handy as a starting point: http://www.cgl.ucsf.edu/home/scooter/CSB/Assorted_Cytoscape_Plugins_Handout.pdf
The original answer may have been to obscure (was that why it was voted down?). Cytoscape is not only useful for visualisation of the resulting network it can also assist you in the actual calculations: In that respect the ExpressionCorrelation plugin would probably be the most useful: http://www.baderlab.org/Software/ExpressionCorrelation
BTW it would be nice if people explain why they vote things down. If an answer is wrong the one taking the trouble to answer in the first place would at least learn what is wrong.
I initially voted the answer down, because it provided a very generic answer. It was just a link to Cytoscape's plugins. Now, with the extra information about a specific plugin and installation instructions provided by the linked page, the answer seems much better. So, I have removed my initial down-vote.
@ CHirs Evelo I want to find out Pair wise corealtion between a list of lncRNA and mRNA can i and build interaction netwrok can i sue cytoscap for it ? if yes which app and how i used 4 ExpressionCorrelation Makes a similarity network where nodes are genes, and edges denote highly
I recommend the use of more refined methods than Pearson correlation. At least, working on partial correlations matrix would be a first step. That involves a bit more of labour, especially because you have to work with positive definite matrix (which mainly means columns and rows should not be linear combinations between others). One way is to shrink your matrix first to correct sur-estimation of correlation coeff biases. Korbinian Strimmer made an amazing work on those topics. You can also try the Bayesian networks, but you would need even more investment to manage those.
You can construct co-expression networks from your gene expression experiments with a variation of Pearson correlation and/or Mutual Information methods with InSyBio BioNets (a demo is available here http://demo.insybio.com and you can request for a free one month access at info@insybio.com). In specific, what makes the difference is the threshold over which you consider that two genes are correlated. The assignment of an arbitrary value to such a threshold, or defining an arbitrary p-value for adding an edge in a co-expression network may lead to over- and under-estimating correlation as already mentioned. With InSyBio BioNets approach this threshold is estimated systematically per gene assuming that its correlation values with the expressions of other genes follow normal distribution. For more details read the relevant white paper: http://insybio.com/pages/bionets
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
Another comment on Pearson Correlation Coefficient. Be very cautious with this metric as it is known to have serious issues with over- and under-estimating correlation because of the influence of small numbers of outliers. It might be a better idea to use Spearman correlation or Euclidian distance.
Speaking of Pearson Correlation Coefficient in relation to gene coexpression, you might also want to check out Mixed-Model Coexpression that was introduced in this publication: "Mixed-model coexpression: calculating gene coexpression while accounting for expression heterogeneity". An R implementation is available on their web-site: http://genetics.cs.ucla.edu/mmc/