Correlation detection strategies in microbial data sets vary widely in sensitivity and precision
This article (which I co-authored)
presents an evaluation of different microbial network inference tools and in
summary shows that the accuracies for detection of simulated ecological interactions in synthetic data are very low
for all the tools.
Shall we drop network inference altogether? I do not think so. First, the evaluation did not
look at taxon groups responding together to environmental conditions or representing alternative communities.
But the detection of such groups is of biological interest and something that network inference can provide likely
with far greater accuracy than the exact ecological interactions (an evaluation of group detection accuracy is still missing).
Second, the evaluation was carried out on synthetic data only. Although it is likely that an evaluation with
biological data will confirm this result, we do not know the performance of these tools on real-world data.
The only evaluation on biological benchmark data that I am aware of (May 2016) was carried out for CoNet as part of the
TARA Oceans network inference.
Many of these known ecological interactions involved endosymbionts, which are particularly easy to detect
and might explain why the sensitivity of CoNet in that case was higher than on the simulated ecological interactions.
The message of this evaluation article is that one should not interpret edges in inferred networks
exclusively as ecological interactions. They may represent ecological interactions in some cases,
but in many cases they represent a common response of a group of microorganisms to environmental
factors (an example is discussed in my tutorial on network construction)
or alternative communities. And of course, there are always many false positive edges, which is the reason why
network inference has to be run with stringent settings
and why resulting networks have to be interpreted with care.