Minimum-Risk Profiles of Protein Families Based on Statistical Decision Theory

Statistical decision theory provides a principled way to estimate amino acid frequencies in conserved positions of a protein family. The goal is to minimize the risk function, or the expected squared-error distance between the estimates and the true population frequencies. The minimum-risk estimates are obtained by adding an optimal number of pseudocounts to the observed data. Two formulas are presented, one for pseudocounts based on marginal amino acid frequencies and one for pseudocounts based on the observed data. Experimental results show that profiles constructed using minimal-risk estimates are more discriminating than those constructed using existing methods.