statistics method to find the relation between multiple variables of 50 samples

Question

[maria2019]

Hi,

I have 50 samples and each sample has 3 variables (example data structure is as below)

     Samples   Var1  Var2  Var3
      S1       10     5     2
      S2       1      4     6
      S2       5      4     0
      .        .      .     .
      S50      2      1     10

What is the right statistics method to find the relation between theses variable and find out if there is a pattern in these variable?

[Asaf]

Other than pairwise correlation?

[maria2019]

yes, I thought there might be some other new methods to work on it

[Alex Reynolds]

Perhaps also look into PCA: https://en.m.wikipedia.org/wiki/Principal_component_analysis

[maria2019]

I am mostly interested to see, say if variable 1 increases, what happens at variable 2, etc. I know that PCA can show the separation between the groups but I need some more detailed method

[German.M.Demidov]

General simple test then - Spearman correlation, however, your distributions seems to be zero inflated and with many ties. Glm may help, but then you need to do plotting.

[n,n]

what exactly are these variables? maybe gene expression?

[maria2019]

I actually am looking RNA-seq (FPKM (log2)), ATAC-seq data (Normalized RPKM ), and methylation % at 50 genes from one sample. Samples are my gene names and variables are RNA-seq, ATAC-seq, and methylation data. I wanna see the relation between these three data at these genes.

[Asaf]

So you want to see if there is a linear model gene ~ ATAC + methylation? Just put the data into a data.frame and use lm()

[maria2019]

Thank you very much! I just tried it

One question, I also tried cor(mydata). I am gonna read through this more in depth but do you think I should get the same result for cor and lm? (mine are different when I try these data)

[Asaf]

See here about comparison of lm to cor: https://lindeloev.github.io/tests-as-linear/

[German.M.Demidov]

The first method is plotting. Each variable separately and one against another. In r you should use functions plot(density(var)) and plot(var1, var2)

[maria2019]

I will try it thanks

[German.M.Demidov]

It may sound obvious, but it does worth time spent 100 percents - linear correlation is good only for some particular data distributions and only plotting may give you an intuition what is correct choice of the test (eg https://en.m.wikipedia.org/wiki/Anscombe%27s_quartet