Differential expression analysis with global network adjustment
1 Department of Epidemiology and Biostatistics, University of Texas Health Science Center San Antonio, San Antonio, Texas, USA
2 Department of Biostatistics, University of North Carolina School of Public Health, Chapel Hill, North Carolina, USA
3 School of Mathematics and Statistics, University of Glasgow, Glasgow, Scotland, UK
4 Department of Statistics, University of Connecticut, Storrs, Connecticut, USA
5 Department of Pediatrics, University of Texas Health Science Center San Antonio, San Antonio, Texas, USA
BMC Bioinformatics 2013, 14:258 doi:10.1186/1471-2105-14-258Published: 23 August 2013
Large-scale chromosomal deletions or other non-specific perturbations of the transcriptome can alter the expression of hundreds or thousands of genes, and it is of biological interest to understand which genes are most profoundly affected. We present a method for predicting a gene’s expression as a function of other genes thereby accounting for the effect of transcriptional regulation that confounds the identification of genes differentially expressed relative to a regulatory network. The challenge in constructing such models is that the number of possible regulator transcripts within a global network is on the order of thousands, and the number of biological samples is typically on the order of 10. Nevertheless, there are large gene expression databases that can be used to construct networks that could be helpful in modeling transcriptional regulation in smaller experiments.
We demonstrate a type of penalized regression model that can be estimated from large gene expression databases, and then applied to smaller experiments. The ridge parameter is selected by minimizing the cross-validation error of the predictions in the independent out-sample. This tends to increase the model stability and leads to a much greater degree of parameter shrinkage, but the resulting biased estimation is mitigated by a second round of regression. Nevertheless, the proposed computationally efficient “over-shrinkage” method outperforms previously used LASSO-based techniques. In two independent datasets, we find that the median proportion of explained variability in expression is approximately 25%, and this results in a substantial increase in the signal-to-noise ratio allowing more powerful inferences on differential gene expression leading to biologically intuitive findings. We also show that a large proportion of gene dependencies are conditional on the biological state, which would be impossible with standard differential expression methods.
By adjusting for the effects of the global network on individual genes, both the sensitivity and reliability of differential expression measures are greatly improved.