Protein alignments are commonly characterized by the probability vectors over amino acids in each column of the alignment. This paper develops various models for the probability distribution of these probability vectors. First a simple Dirichlet distribution is used, then a mixture of Dirichlets. Finally a componential model employing a `density network' is described. These models are optimized and compared using Bayesian methods.
@UNPUBLISHED{MacKay94:amino,
AUTHOR ="D. J. C. MacKay",
TITLE ="Models for Dice Factories and Amino Acid Probability Vectors",
YEAR ="1995",
NOTE ="Unpublished",
ANNOTE ="Date submitted: ; Date accepted: ; Collaborating institutes:
MRC Laboratory of Molecular Biology, Cambridge"}