Analyzing correlation of non normal distributed data from Sequencing analyses

[#1]

Hello,

I may have not dig enough but I don't find the answer in the community.

We have NGS like data, expression of genes, that we would like to correlates between different sample zones. The data are not normally distributed. 

We have various inputs saying that we should use log transformation to get a normal distribution. 

We want to study tha correlation, can we do it on non-normal distributed data? Do we need to go for the log distribution? 

The correlation are not the same using one or the other method so we want to be sure we are doing it the right way are we are not statisticians.

Thanks to anyone that can help us on that point.

Have nice day

Marie

[#2 · Accepted solution]

Hi @mgerusdurand , thank you for your question!

For your data, is it sparse (a lot of zeros)? If the data is not very sparse, then generally, like RNA-seq data, we would recommend transforming the data, e.g. log-transform, and then perform correlation analysis, dimension reduction and visualization (PCA, UMAP, TSNE),  and clustering.

There are also different types of correlations -- pearson correlation, spearman's correlation, kendall's tau, etc. Depending on what your research interest is, you may choose different methods. 

[#3]

Thanks @MeichenDong !

We don't have a lot of zero, these are not tandard RNA seq data because they are spatially resolved. I don't want to "hide" outliers by a log transformation so I don't get the recommended transformation. Pearson correlation on non parametric data sounds like answering our question in this particular setting.

Thanks for your help

[#4]

I am also having the same issue.

[#5]

Hi @CrossoverBird55 , is there any specific problem that we could help?

The way to transform / normalized data also depends on the downstream analysis plans.