Similarity matrix to distance matrix for protein sequences

Question

[kbaitsi]

I have used R to calculate a similarity matrix for 11 proteins (histones) from a fasta file. Then I need to turn the similarity matrix into a distance matrix in order to use it in hclust. I have used sim2dist and also dist with all methods (euclidean, maximum, manhattan, canberra, binary, minkowski). I have excluded the binary method but I am not sure which is the best way to calculate the distance from the rest of my options. Any thoughts?

[Jean-Karim Heriche]

There are a few common and generic ways of turning a similarity into a distance such as:

• d = max(s) - s (e.g. if similarity is cosine then max(s) = 1)
• d = 1/(s+1)
• d = exp(- s^a) with a being a parameter In fact, any function that is strictly decreasing will do.

[kbaitsi]

Thank you for your answer, sim2dist does what you wrote in the first bullet. I was just wondering if there is a preferable way when it comes to protein sequences or it doesn't matter?

[Jean-Karim Heriche]

What matters most is the choice of the original measure of similarity. It has to capture the notion of proximity/similarity that is relevant to the question you're trying to address. When converting you need to make sure that distribution properties that are important for the clustering are preserved.

[kbaitsi]

I have used the pairwise alignment function and a blosum subtitution matrix. Thanks a lot for your time and answer.