Hi, I am wondering how the correlation in co-expression network is calculated. I mean, it would be expected to have both positive and negative correlation.

Created by Andras Hartmann ahartmann
Hi We considered two possibilities for generating co-expression networks: (1) taking the top k edges with the highest absolute value of the correlation coefficient or (2) taking the top k edges with the strongest positive correlation. We used the second option because in our experience negative correlations are less informative than positive correlations for constructing networks, and we see little evidence for negative correlations between TF-target gene pairs in practice. For example, in the DREAM5 network inference challenge, the second approach obtains far better performance than the first approach. Indeed, also when looking at the set of established TF-target gene interactions (the gold standard) for E. coli and yeast, we found that the distribution of correlation coefficients was clearly shifted towards positive correlation compared to non-interacting pairs, but we did not see evidence for strong negative correlations between TF-target pairs (Figure S8, [Marbach et al. 2012](http://www.ncbi.nlm.nih.gov/pubmed/22796662)). (In contrast, for the simulated network, there was excess positive and negative correlation for truly interacting pairs as we built the model based on the assumption that both positive correlation (activators) and negative correlation (inhibitors) are important in transcriptional networks). My intuition is that activation / positive correlation dominates in transcriptional networks, which corresponds to a general mode where genes are turned off by default and activated when needed. In contrast, for strong negative correlation a gene would have to be activated by default and inhibited when not needed, which is less efficient. There is of course inhibition in transcriptional regulation, but my intuition is that it generally doesn't lead to a signature of strong negative correlation across conditions in expression data. Independently, one of the groups providing a correlation-based network also considered both options and found that focusing just on the positive correlations gave better results. Best Daniel

Negative correlation in Co-expression network page is loading…