In edgeR, one usually estimates a common, trended and tag wise dispersion in order to estimate the mean-variance trend over all genes.

I have noisy data with large BCVs. It is actually single-cell RNA-seq, and just for comparison, I would like to analyse it with edgeR. However, the common, trended and tag wise dispersion estimates are bounded. One can fix the bounded common dispersion with the 'interval' argument, but not the other two.

The estimateDisp function does not have these bounded estimates, but does not allow the inclusion of weights, which the other three functions above do.

Thus I am interested in an edgeR workflow which

(a) is not bounded in its dispersion estimates, and

(b) allows the inclusion of weights.

Is there any way I can relatively easy tweak the code to do this?

Thanks

I'm not sure what you mean by "bounded" dispersion estimates. All functions (including estimateDisp) use a grid search on the adjusted profile likelihood to estimate the NB dispersion, so the only bound is on the range of the grid. For estimateDisp, for example, the grid spans dispersion values from ~0.0001 to over 100 by default. The presence of an upper bound would only be a concern if you have NB dispersions over 100, and if that's the case, the variability is so high that I'd struggle to see what information that gene could provide for detecting DE. Even for noisy single-cell RNA-seq data, values of 1-10 seem to be more typical.

As for the weights, estimateDisp can be easily modified to support weights in the calls to adjustedProfileLik and glmFit. However, the sensibility of the weighting scheme would then become your responsibility. If I remember correctly, a larger weight for a count will scale down its variance and increase its influence on the GLM fit. Thus, your weights should probably be computed based on some measure of the variance, otherwise it wouldn't make much quantitative sense to scale the variances by some other unit of measurement.