A correction for this article has been published in Biology Direct 2009, 4:10

Research

Pitfalls of the most commonly used models of context dependent substitution

Helen Lindsay¹ , Von Bing Yap² , Hua Ying¹ and Gavin A Huttley¹

¹  Computational Genomics Laboratory, John Curtin School of Medical Research, The Australian National University, Canberra, ACT 0200, Australia

²  Department of Statistics and Applied Probability, National University of Singapore, Kent Ridge, Singapore

author email corresponding author email

Biology Direct 2008, 3:52doi:10.1186/1745-6150-3-52

Published:	16 December 2008

Abstract

Background

Neighboring nucleotides exert a striking influence on mutation, with the hypermutability of CpG dinucleotides in many genomes being an exemplar. Among the approaches employed to measure the relative importance of sequence neighbors on molecular evolution have been continuous-time Markov process models for substitutions that treat sequences as a series of independent tuples. The most widely used examples are the codon substitution models. We evaluated the suitability of derivatives of the nucleotide frequency weighted (hereafter NF) and tuple frequency weighted (hereafter TF) models for measuring sequence context dependent substitution. Critical properties we address are their relationships to an independent nucleotide process and the robustness of parameter estimation to changes in sequence composition. We then consider the impact on inference concerning dinucleotide substitution processes from application of these two forms to intron sequence alignments from primates.

Results

We prove that the NF form always nests the independent nucleotide process and that this is not true for the TF form. As a consequence, using TF to study context effects can be misleading, which is shown by both theoretical calculations and simulations. We describe a simple example where a context parameter estimated under TF is confounded with composition terms unless all sequence states are equi-frequent. We illustrate this for the dinucleotide case by simulation under a nucleotide model, showing that the TF form identifies a CpG effect when none exists. Our analysis of primate introns revealed that the effect of nucleotide neighbors is over-estimated under TF compared with NF. Parameter estimates for a number of contexts are also strikingly discordant between the two model forms.

Conclusion

Our results establish that the NF form should be used for analysis of independent-tuple context dependent processes. Although neighboring effects in general are still important, prominent influences such as the elevated CpG transversion rate previously identified using the TF form are an artifact. Our results further suggest as few as 5 parameters may account for ~85% of neighboring nucleotide influence.

Reviewers

This article was reviewed by Dr Rob Knight, Dr Josh Cherry (nominated by Dr David Lipman) and Dr Stephen Altschul (nominated by Dr David Lipman).