You got completely different results depending on whether you compared logFC or linear data? Were you comparing like with like in both cases?

I.e. if your overexpression is listed as logFC, compare that with logFC, while if your overexpression is expressed linearly, compare that with linear.

If your correlation is good (or indeed, if there IS a correlation) it shouldn't really matter which you use. If your correlation is weak and your data highly variable (i.e. some fold changes are 1.4-fold, others are 50-fold) then you may find linear comparisons give you spurious correlations purely because you have a clump of data with one outlying point (effectively therefore being a 'curve' drawn between two points), whereas log xformed data will account for this and correctly show no correlation. Alternatively, if your correlation is modest and your FC values are also modest, then log-xforming everything will simply clump the data together needlessly. It is difficult to advise further without knowing your genes of interest or your biological context. For example, some transcription factors are essentially absent under condition X, but expressed under condition Y, so here a fold-change would be a useful metric (you may be seeing changes of 20- or even 100-fold, from 'basically none' to 'some'), whereas other genes (even other transcription factors) might be expressed at quite high levels constitutively, but are further boosted by 50-100% (i.e. 1.5 or 2-fold) under specific conditions: this is a much more modest increase in fold-terms, but a massive increase in actual transcript, so linear comparisons might be more appropriate.

Or, perhaps, you don't have a correlation.

Be wary of comparing many different genes: you will get apparent correlations by chance quite easily this way. Maybe use a more sophisticated method to see if en-bloc correlations exist (like PCA).

Thank you for your reply!

The data was in the same state - my genes of interest's expression values and the overexpressed one are in one data frame and both are LogFC values.

The number of genes is about 60. LogFC's of genes are ~ -3...+3, LogFC of HE/Knockout ~ -4...+6.

The aim of the calculation was to find out whether the difference in transcription factors expression levels is correlated with the level of gene, plants were transformed with, expression.

I performed the correlation test using

rcorr(MyData, type=c("pearson","spearman"))

of R package Hmisc.

And got the result consist of a few genes, that have reliable correlation to the OverExp. gene expression level (p.values < 0.05).

Than I thought that maybe if I'm trying to conduct a linear correlation analysis, than I need to transform MyData to linear one. So I did it and performed analysis again.

In the result I get a few genes reliably correlated to my OverExp. gene, but it was another genes, with no matches to the previous calculations.

I thought that the second method may be better, but can't find neither the confirmation nor denial of it.

Log2 or log10?

I am assuming log2, in which case 6 is substantial but biologically plausible (~64 fold) rather than stupidly excessive (100000 fold).

Also, does your R package measure both pearson and spearman correlations? That's implied by the text above, but this is not advisable.

The former test is for normally distributed data, the latter for non-normal: use the test that is appropriate to your data (which may differ depending on whether you're comparing linear or log xformed data).

At the end of the day, without looking at your data, I cannot really advise further, but I would recommend you actually look at the correlations your analysis is spitting out (i.e. manually inspect the datapoints) and see if you believe them or not. I would also seriously consider putting the threshold for significance at much lower than 0.05, because with 60 genes, ~three will be significant by your current threshold even if there is no correlation.

The data is Log2 transformed.

So, would it be ok, if I check the distribution of my log2-transformed data and than perform proper (depending on distribution) type of correlation analysis, considering that the correlation is reliable if p.value is less than 0.01?