Phred score cutoff vs. sequence length for an unbiased comparison between different cohorts


I am studying the effect of a certain treatment on the gut microbial composition of different organisms which vary in sequencing quality considering Phred score plots. In order to have an unbiased pipeline, I was wondering if I should choose a specific Phred score for all OR have the same trim and truncate parameters so that the resulting sequences would have comparable lengths?

The same issue exists regarding sequencing depths which vary between the studied organisms. In this case, should the rarefaction depth be specific to each studied animal? Would that make sense to compare the results (for example clustering patterns in each case) or would it be kind of like comparing apples and oranges?

Thanks in advance for your inputs! :slight_smile:

Hi @Parix,
Great questions!

I am assuming these are from different sequencing runs?

The most important thing is that the sequence lengths are the same if you plan to use denoising methods to denoise and dereplicate. If sequences are not the same length, each run will have its own unique ASVs and you would have trouble comparing them.

So if the reads are single-end it will be important to trim them at the same lengths. If they are paired end, then theoretically the merged reads should yield the same length and same amplicons.

Another way to handle this if you must trim to different lengths is to use closed-reference OTU clustering to cluster the reads against known full-length reference sequences. It’s a lazy but efficient way to solve what is sometimes an intractable problem.

You should rarefy all to the same depth if you plan to compare these for alpha or beta diversity analyses. Otherwise it would be like :apple:s and :orange_circle:s!

Good luck!


@Nicholas_Bokulich, Thanks for your answer :slight_smile:

Yes, they are from different runs.

Is it ok for the paired-end reads to have comparable lengths? since it is not convenient to have the exact same length across different cohorts.

Actually, I am not comparing them directly. Rather, I want to see for example if the samples from the same treatment cluster together in different cohorts. Since the sequencing depth varies a lot in different cohorts (orders of magnitude), I thought I would lose a lot of data if I chose the shallowest depth for all. Wouldn’t that make sense then to rarefy separately for each cohort according to its specific sequencing depth and then compare if the samples “behave” similarly regarding the treatment?

If you want to compare these directly, you will need to:

  1. use the exact same length and site if you want to compare ASVs, otherwise you are introducing bias
  2. compare using the taxonomy instead of ASVs (in which case there is still bias, but it is reduced somewhat if the regions are similar)
  3. use closed-reference OTU clustering as described above

You are correct, the normalization (in this case rarefaction) can be determined on a case-by-case basis — but rarefaction or other normalization is required to be at the same depth for each comparison that does occur. But rarefying at different depths for different comparisons is fine.

1 Like

This topic was automatically closed 31 days after the last reply. New replies are no longer allowed.