Fig. 1

Scaling normalization approaches from the perspective of compositional correction. a Features S and A have similar absolute abundances in two experimental conditions, while B has increased in its absolute abundance in condition g due to technical/biological reasons. Because of the proportional nature of sequencing, increase in B leads to reduced read generation from others (compositional bias). An analyst would reason A and S to be significantly reduced in abundance, while, in reality they did not. b Knowing S is expressed at the same concentration in both conditions allows us to scale by its abundance, resolving the problem. DESeq and TMM, by exploiting rerefence strategies across feature count data (described below), approximate such a procedure, while techniques that are based only on library size alone like RPKM and rarefication/subsampling can lead to unbiased inference only under very restrictive conditions. Approaches available for sparse settings are indicated. Wrench is the proposed technique in this paper