pearson correlation or Euclidean distance for clustering?

Question

[mrashad]

I have a matrix of multi omics expression and need to make a clustering using Hierarchical clustering and k means but confused between the used distance Euclidean distance or Pearson correlation.

Is there any guide for which one of them should be used in expression data?

Answer 1

[Kevin Blighe]

There is neither a guide nor standard for this.

If using either of Euclidean distance or Pearson correlation, your data should follow a Gaussian / normal (parametric) distribution. So, if coming from a microarray, anything from RMA normalisation is fine, whereas, if coming from RNA-seq, any data deriving from a transformed normalised count metric should be fine, such as variance-stabilised, regularised log, or log CPM expression levels.

If you are performing clustering on non-normal data, like 'normalised' [non-transformed] RNA-seq counts, FPKM expression units, etc., then use Spearman correlation (non-parametric).

As usual, get intimate with your data, know its distribution, and thereafter choose the appropriate method(s).

Kevin

[Hamid Ghaedi]

A good point to rise is data distribution importance for choosing distance measures in clustering analysis. Thanks This is my understanding of differences between Euclidean distance or Pearson correlation distances application for gene expression clustering: When we are interested in considering overall expression profiles (up and down), correlation-based measures (i.e. Pearson correlation) would be of choice. In other cases, we may want to cluster observations with the same magnitude of dysregulation together. In this way observations with high value of features would cluster together. In these cases, Euclidean distance would be our choice for dissimilarity matrix calculation.

[mrashad]

Thanks a lot I got it

[mrashad]

I got it, thanks a lot for this fruitful answer.