5 months ago by
Without replicates it is pretty much meaningless to do differential analysis as it is impossible to find the variability for each gene. In a child-vs-parent comparison there is some ability to measure variance, because you have both mother and father, and conceptually you can measure the mother-father variance and then see where the child lies in relation to that. Both edgeR and voom (from raw read counts) will allow you to do this sort of experiment. Limma-trend will allow you to input CPMs I think, but you are another step removed from ideal then.
Be aware however of the following:
- This is making the assumption that expression variation is the same in the child as it is in the parents. This may not be true
- Dispersion estimates made using only 2 samples are going to be poor. This will lead to both poor power and poor accuracy. In theory this should be accounted for in the FDR value, but in practice this is likely not the case.
- Low power means more false positives.
Thus this analysis may give you some signal amongst the noise, and may even be enough to give you enrichments for downstream analyses, but I would not trust the result for any particular gene without further evidence.
An alternative approach might be to calculate the variance for each gene, and then de-trend it (effectively calculate the dispersion). If you then assume that the dispersions should be normally distributed, you could calculate a Z score for each gene compared to the mean and variance of all genes, thus identifying genes that are outliers.
This would have the advantage that you would find genes where any of the individuals could be the outlier (thus cases where child is more like mother or father as well as cases where the child is the outlier). It has the disadvantage that it would also identify cases where expression of a given gene is just naturally more variable than others.