There is powerpoint file from hosted at the server of Department of Mathematics at the University of Michigan which states

PAM (Point Accepted Mutation) uses sequence data for proteins which can be aligned with 99% identical sequence. In such aligned sequences, one counts the frequency of each residue, and of each pair of residues in a "column".

Recall: using p(a,b) = prob. that a (amino acid residue) mutated into b.

Basically, the score s(a,b) from this data is a "LOD score":

   s(a,b) = log(p(a,b)/p(a)p(b)).

Comparison of two "models": significant association vs. independent randomness.