Finding remote protein homologs with hidden Markov models

Detecting remote homologs by sequence similarity gets increasingly difficult as the percentage of identical residues decreases. The aim of this work was to investigate if the performance of hidden Markov models could be improved by ignoring the subsequences that exhibit high variability, and only concentrate on the truly conserved regions. This is based on the underlying assumption that these high variability regions could be unnecessary, or even misleading, during search of remote protein homologs. In this paper we challenge this assumption by identifying the high and low variability regions of multiple alignments and modifying models by focusing them on the conserved regions. The high variability regions are located with information theoretic measures and modeled by free insertion modules, which are special nodes that can be used to model arbitrarily long subsequences with a uniform probability distribution. The results do not support a definitive conclusion since a few cases exhibit a performance increase, while the general trend is that the performance decreases when ignoring high variability regions. Two supplementary tests suggest that when there is a significant performance loss due to deletion of high variability nodes, a much smaller decrease occurs when the nodes are preserved but the position-specific amino acid distributions are removed. Taken together, these results support the hypothesis that there is some valuable information present in the high variability regions that enable the model to better discriminate between true and false homologs; and that other constructs for the high variability regions could perform better.