...

One important usage of the shrinkage plot is a meaningful setting of low expression threshold in Low expression filter section (Figure 6). For features with low expression, the proportion of zero counts is high. Such features are less likely to be of interest in the study, and, in any case, they cannot be modeled well by a continuous distribution, such as Lognormal. Note that adding a positive offset to get rid of zeros does not help because that does not affect the error term of a lognormal model much. A high proportion of zeros can ultimately result in a drop in the trend in the leftmost part of the shrinkage plot (Figure 5).

...

Figure 10 contains a pie chart for the dataset whose shrinkage plot is displayed in Figure 4. Because of a small sample size (two groups with four observations each) we see that, overall, shrinkage is beneficial: for an "average" feature, Akaike weight for feature-specific Lognormal is 25%, whereas Lognormal with shrinkage weighs 75%.

Using multimodel inference appears to be a better alternative to the ad hoc method in DESeq2 that switches shrinkage on and off all the way. Once again, it is both technically possible and emotionally tempting to automate the handling of abnormal features by enabling both Lognormal models in GSA and applying them to all of the transcripts. Unfortunately, that can make the results less reproducible overall, even though it is likely to yield more accurate conclusions about the drastically outlying features.

...