You could try to obtain normal / healthy RNA-seq samples by searching GEO, SRA, et cetera. However, my preference would be to process the data as normal (normalisation and transformation), and to then convert the normalised+transformed expression levels to Z-scores. Then, any gene that has Z > 1.96 or Z < -1.96 will be statistically significantly (p<0.05) increased or decreased in expression.

Kevin

The question you are interested in here, is whether each somatic mutation has an effect on the expression of the gene in question. Thus, in each case you DO have two groups of samples - wildtype samples, and mutant samples. The fact they are all cancer is immaterial.

Thus, what you are doing here is an eQTL study, rather than a differential expression study. Unfortunately 5 samples is a very small number of samples to do this with. The ideal would be to fit a limma/edgeR/deseq model for each mutation, dividing into those that have the mutation and those that don't, and check if the gene in question is differential. But if your list of mutations is more than a few, then this is going to get very tiresome, and unless you do some trickery with recalculating the mutliple testing correction, the multiple testing is going to absolutely wipe you out.

Alternatively you could perform a variance stablaising transform, like VST, rLog or (i think) voom, and then for each somatic mutation, take the estimates for that gene, divide to mutant and wildtype samples, and do a t-test.

Finally, if each mutation only occurs in one sample (which I guess is fairly likely), do something like what @Kevin Blighe suggested, but use the other 4 samples to calculate mean and SD on variance stabilized estimates. Although, again, you might struggle with only 5 samples.