Why you should think twice before rarefying your 16S data

According to a recent paper, the common practice of rarefying (randomly subsampling your 16S reads), is statistically incorrect and should be abandoned in favor of more sophisticated ‘mixture-model’ approaches favored by RNA-Seq analysis software.

The authors give a basic thought experiment to help clarify their reasoning.   I’ve copied the figure below.  It basically shows what happens when one rarifies library “B” from 1000 reads to 100 reads in order to match library “A.”  This is a standard procedure, even in the QIIME workflow.  By dividing the library size by 10, the variance of the data in B goes up by 10, and thus the statistical power to differentiate between OTU1 and OTU2’s abundances in samples A and B is lost.

Fig 1. An example of the effect of rarefying on statistical power.
Fig 1. An example of the effect of rarefying on statistical power.

As the authors point out in the paper:

“Although rarefying does equalize variances, it does so only by inflating the variances in all samples to the largest (worst) value among them at the cost of discriminating power (increased uncertainty). “

The solution to the rarefying problem is to take advantage of methods from the RNA-Seq world in order to determine differential taxonomic abundances, just as differential transcript abundances are calculated by the same methods.   The R package, phyloseq, provides an R container for taxonomic datasets from Qiime (in the BIOM format).  It also provides a convenient set of extensions that allow analysis of this data with RNA-Seq methods like edgeR and DESeq2.

Should you trim your RNA-Seq reads?

According to a new paper, basically, no.   Actually that is an oversimplification, but the authors find that quality trimming of RNA-Seq reads results in skewed gene expression estimates for up to 10% of genes.   Furthermore, the authors claim that:

“Finally, an analysis of paired RNA-seq/microarray data sets suggests that no or modest trimming results in the most biologically accurate gene expression estimates.”

First, the authors show how aggressive trimming affects mappability in Figure 2:

Rna-Seq reads trimming effects.
Influence of quality-based trimming on mappability of reads.

You can see that as the threshold becomes more severe (approaching 40), the number of RNA-Seq reads remaining drops off considerably, and the overall % mappability increases.  Overall, you’d think this would be a good thing, but it leads to problems as shown in Figure 4 of the paper:

Rna-Seq reads.
Isoform and gene expression levels after trimming.

Here you can see in (a) how increasingly aggressive trimming thresholds lead to increased differential expression estimates between untrimmed and trimmed data (red dots).  Section (b) and (c) also show that the number of biased isoforms and genes, respectively, increases dramatically as one approaches the Q40 threshold.

One way to correct this bias is to introduce length filtering on the quality-trimmed RNA-Seq reads.  In Figure 5, the authors show that this can recover much of the bias in gene expression estimates:

Isoform and gene expression levels after length-filtering.
Isoform and gene expression levels after length-filtering.

Now in (b-d) it is clear that as the length filter increases to 36, the number of biased expression estimates goes rapidly down.   There seems to be a sweet spot around L20, where you get the maximum decrease in bias while keeping as many reads as possible.

Taken together, the authors suggest that aggressive trimming can strongly bias gene expression estimates through the incorrect alignment of short reads that result from quality trimming.  A secondary length filter step can mitigate some of the damage.   In the end, the use of trimming depends on your project type and goals.  If you have tons of reads, some modest trimming and length filtering may not be too destructive.  Similarly, if your data are initially of low quality, trimming may be necessary to recover low-quality reads.  However, you should be restrained in your trimming and look at the resulting length distributions if possible before deciding on quality thresholds for your project.