I have created co-occurrence network based on Spearman's correlations for several times with different projects. My research focuses on microbiome. So the networks were the correlations between different microbial taxa. When I filtered the correlations by coefficients above 0.7 or 0.8, I only got positive correlations (or at least most of them positive correlations). My colleagues also got the same issue as me. After I checked many literatures, they also got like mostly the positive correlations based on filtering with a high cutoff of coefficients.
Is there any reason why this happened? Then how to interpret from all positive correlations?
I would simply make a note of this for future reference as you have already done. A part of the issue may be the range in values that you are working with. If I look at plant growth and yield I will get a positive relationship for nearly every variable if I am growing the plant on nutrient deficient soil. I add a bit of nitrogen and I get more growth and higher yield. I add phosphorus, potassium, boron, copper, manganese, iron, magnesium in small quantities and I get a positive response. It isn't until I start adding the nutrients in excess that I get other relationships. Might this kind of thing be taking place in your system? You keep the variables all within a "reasonable range of values" and as a result you always get one relationship. Experimentally it is often good to have a few treatments that are unreasonable in order to test the system.
Hi Haitao,
I also applied the network analysis in my study for revealing the co-occurrence pattern but between trace elements and genes. And I also got a lot of correlations (also includes negative correlations but mostly positive ones). One of the reviewer recommonded the Benjamini-Hochberg Procedure FDR (to confirm the significance) and randomization procedure check (network analysis on your same data but randomized picked ones to double check the pattern). If you have suspection, you could try them to confirm your results. I have performed it on my network, and found out if the cut-off is high enough, the result remians the same (I used pearson's r of 7.0, and P value of 0.05). Hope it is useful for you. :)
Best,
Yi
You can do some adjustment or randomized check as Zhao recommended, but I suspect if they can change the existence of more positive associations in any case.
If you are doing spatial analysis, I think more positive is not a problem because there might be a lot of structural zero of a given tax in the abundance profile. It is rare to generate negative correlations between taxes unless two taxes are constantly abundant in many locations and one can "kill" another. And usually the absolute number of the weight of such negative correlations is much lower. Structural zero sometimes is good in spatial analysis because it can provide a baseline for different environments, but of course, you can remove them to increase the number of negative correlations. If you are doing temporal analysis, improving the resolution of sampling (to detect time lag precisely), or removing the structural zero will help infer more negative correlations.
- Badges
- Science topic
- Similar topics
- Comment