I am noticing an interesting trend on the dispersion plot where dispersion estimates are increasing for genes with extremely high counts:

These are selection experiments where we expect a high degree of variance for highly-expressed genes (a given gene may be highly expressed in one condition but not another). Still, I think the model fit is generally performing fairly well across the range of counts in this example. Would you be wary of these dispersion estimates leading to unpredictable p-value estimates from the NB GLM?

"we expect a high degree of variance for highly-expressed genes (a given gene may be highly expressed in one condition but not another)"

The dispersion doesn't increase if there are changes across condition, only due to variation within condition.

The red line in dispersion plot seems like a good reflection of the trend of the MLE estimates.

"Would you be wary of these dispersion estimates leading to unpredictable p-value estimates from the NB GLM?"

The dispersion posterior estimates seem to track closely with the trend so I don't see any issue.