Research article

Missing channels in two-colour microarray experiments: Combining single-channel and two-channel data

Andy G Lynch¹ , David E Neal¹ , John D Kelly¹ , Glyn J Burtt¹ and Natalie P Thorne^1,2

¹  Department of Oncology, University of Cambridge, Cambridge, UK

²  Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Cambridge, UK

author email corresponding author email

BMC Bioinformatics 2007, 8:26doi:10.1186/1471-2105-8-26

Published:	25 January 2007

Abstract

Background

There are mechanisms, notably ozone degradation, that can damage a single channel of two-channel microarray experiments. Resulting analyses therefore often choose between the unacceptable inclusion of poor quality data or the unpalatable exclusion of some (possibly a lot of) good quality data along with the bad. Two such approaches would be a single channel analysis using some of the data from all of the arrays, and an analysis of all of the data, but only from unaffected arrays. In this paper we examine a 'combined' approach to the analysis of such affected experiments that uses all of the unaffected data.

Results

A simulation experiment shows that while a single channel analysis performs relatively well when the majority of arrays are affected, and excluding affected arrays performs relatively well when few arrays are affected (as would be expected in both cases), the combined approach out-performs both. There are benefits to actively estimating the key-parameter of the approach, but whether these compensate for the increased computational cost and complexity over just setting that parameter to take a fixed value is not clear. Inclusion of ozone-affected data results in poor performance, with a clear spatial effect in the damage being apparent.

Conclusion

There is no need to exclude unaffected data in order to remove those which are damaged. The combined approach discussed here is shown to out-perform more usual approaches, although it seems that if the damage is limited to very few arrays, or extends to very nearly all, then the benefits will be limited. In other circumstances though, large improvements in performance can be achieved by adopting such an approach.