Reader Comments

Post a new comment on this article

Referee Comments: Referee 2

Posted by PLOS_ONE_Group on 04 Sep 2007 at 10:34 GMT

Reviewer 2's Review

“The paper intends to unravel part of the functional network of miRNAs in brain tumours. As the authors point out, many experiments with similar purposes first alter the endogenous levels of miRNAs, which could perturb the underlying biological process. The authors aim to overcome this limitation by searching for correlations between unperturbed expression profiles of miRNAs and mRNAs.

One would assume that the expression levels of direct mRNA targets should correlate negatively with the levels of the controlling miRNA. Secondary targets could correlate both negatively or positively, depending on what factors lie between. Another source of expected correlation, as the authors mention, are transcripts that contain the miRNA genes, which should therefore show positive correlations. Lastly, one would want to avoid, as much as possible, spurious correlations that occur by chance.

This last point is what brings me to my mayor concern about this paper. Given the authors' conclusion that (page 3, paragraph 3) "the correlations between miRNAs and mRNAs contained information regarding interactions" they should first try to eliminate other factors that could make such a difference between the observed coefficients and the random case. I would urge them to consider that the mRNA expression profiles are highly structured (probes representing the same gene can have similar profiles, different genes can have similar profiles due to similar regulation, etc). This is probably the main reason why they obtain such a "significant" p-value. In fact, the authors do mention (page 10, paragraph 1) that 2,563 (36%) of the probes they are using match splice variants of the same genes and are highly correlated.

As a better control, I would suggest keeping the mRNA profiles intact, and randomizing the miRNA profiles. Even then, some miRNAs could have similar profiles, which should be taken into account in the randomization trial. One possibility would be to cluster the miRNA expression profiles before calculating any correlation.

In fact, it might be interesting to see what results are obtained when the profiles of both mRNA and miRNAs are clustered. The correlation results might even be more robust. Functional enrichment tests (of GO terms for instance) could be performed on the clusters. The statistical analysis should also benefit from only having to calculate correlations between clustered profiles by reducing the number of expected false positives. I would even argue that, if no significant correlations are found when using the clustered data, then the observed correlations in un-clustered data are probably spurious.”

N.B. These are the general comments made by the reviewer when reviewing this paper in light of which the manuscript was revised. Specific points addressed during revision of the paper are not shown.