How could we compare the "quality" of different probabilistic models for sequence alignment?Lets consider the simplest case of a Needleman-Wunsch or Smith-Waterman algorithm using a substitution matrix and a gap function. I learned about some different substitution matrices in recent bioinformatics classes and was wondering how we could empirically determine which of those are "better". Some of the substitution matrices were derived from theoretical knowledge, others like BLOSUM and PAM are derived using probability theory from some example alignments.Obviously, the "best" matrix depends on the precise purpose you want to use it for. But lets assume we have some example alignments which are representative for the intended purpose of the user. Then how could we evaluate which method for building a probabilistic matrix is better?
I didn't find the right solution from the Internet.