Distances: my personal experience
Just a few suggestions according to my own experience. For a more formal and detailed treatment go to Wikipedia or a maths manual.
For a very rich directory on string similarity go to this link
Quantitative Data Arrays
For quantitative data (e.g. transcription signals from microarray experiments) you can use
Euclidean distance, which is D = √ ((x1 - x2)2 + (y11 - y2)2 + (z1 - z2) + (...)2 + ...)
it is the geometric distance you are used to calculate in 2 dimensions, just extended over N-dimensions
be careful, I'd suggest to use it only when:you are sure that the data are in the same magnitude scale
you do not want to consider anti-correlated patterns similar
Correlation-based
this distance is ideal when you want to group objects with inter-dependent behavioral trend
for inst, for D = 1 - Correlation Index => similar = whenever A goes up, always B goes up as well; dissimilar = whenever A goes up, B always goes down
whereas, for D = 1 - Abs (Correlation Index) => similar = whenever A goes up, always B goes up as well, AND whenever A goes up, B always goes down
for these reason, I usually prefer Pearson-based distance when clustering genes by expression signals (Affymetrix arrays)
Binary Arrays
Hamming = number of discordant array elements (e.g. 1010 vs 1100 => D = 2)
this distance was developed to measure transmission errors in binary coded strings
Jaccard-based = 1 - (number of concordant 1s in the two arrays)/(number of positions with 1 in either of the two arrays)
(e.g. 1010 vs 1100 => D = 1 - 1/3 = 2/3)
this distance is useful when array positions correspond to elements which can belong (value = 1) or not belong (value = 0) to the set associated to the array
(e.g. given a list of Transcription Factors, you can associate to every gene a binary array to express whether they are regulated by each TF or not)
(e.g. given a list of interaction partners, you can associate to every protein a binary array to express whether they are interacting with each partner or not)
Hamming distance is not good, in my opinion, when the strings compared have a very unequal 1/0 content, and the meaning of 1s and 0s is related to set-membership (as in the examples above).
E.g., consider these strings, and the choice made by Hamming and Jaccard:
11 000 000
is more similar to 10 000 001 according to Hamming (D,,H,, = 2, D,,J,, = 2/3)
is more similar to 10 000 001 according to Jaccard (D,,H,, = 3, D,,J,, = 3/5)
110 100 000
is more similar to 011 000 000 according to Hamming (D,,H,, = 3, D,,J,, = 3/4)
is more similar to 110 111 101 according to Jaccard (D,,H,, = 4, D,,J,, = 4/7)
Mapping the first problem to interactions, let's say
A1 interacts with B, C
A2 interacts with B, D
A3 interacts with B, C, E, F, G
is A1 neighborhood more similar to A2 (Hamming's choice) or A3 (Jaccard's choice)
In addition, neither of these two distances is suited when some array positions are noisy, whereas others are information-rich, as in the following fictional examples:
110 010 001
110 100 000
110 000 110
000 101 100
001 101 011
000 111 000
clearly, 1st and 2nd position are always co-conserved, as well as 4th and 6th; the co-conservation of other residues is irregular; therefore, we may think that the co-conserved positions hold more information than the other ones.