I have hundreds of samples with different sample sizes/counts (3,000 to 150,000). Due to the uneven sizes, comparing the richness between samples can be tricky without rarefying. However, if I rarefy the data to 3,000 counts per sample, many samples' richness would drop significantly, which is also tricky. I wonder if I should use the rarefied richness or the normal richness for data interpretation.
I personally don't like rarefying, which might throw away many rare taxa. I know Shannon which takes into account the relative abundance of each species can to some extent solve this problem. But what if I'm specifically interested in the richness?
I hope someone could provide some enlightening comments on this.
You might check this out:
Article Waste Not, Want Not: Why Rarefying Microbiome Data Is Inadmissible
You just have to make the effort to calculate the accurate diversity
Haitao Wang, don't use 'rarefy' (random subsampling without replacement). If you habe read the paper that Khondoker Dastogeer shared, you will find out that it is "statistically inadmissible". With rarefy, you will receive biased libraries because of the random OTU picking. There are many other tools available for normalising microbiome data, which is crucial for alpha diversity analysis. We recently published an algorithm (SRS) that can be used as an alternative to rarefy and compared the two:
Article Improved normalization of species count data in ecology by s...
The algorithm is implemented in R and easy to use.
Feel free to sent me a message or reply to this answer if you have questions or comments.
Much success with your data, Lukas
Hi Lukas Beule I tried to apply your SRS method to my seq data, but face a problem, which I think is the same Haitao struggles with, which is that I have to scale everything down to my smallest sample - and thereby thronging away A LOT of perfectly good data.
In my case I have sequence data ranging from 1,500 to 20,000 reads per samples. I would have to set Cmin to 1,500....
Do you have any comments/suggestions to that, is it better to discard more samples to begin with, i.e. samples with a low total read count...? but than you loose samples and get potentially unbalanced sample representation from your study design/experiment.
Hope someone will comment on this.
Best, Mathilde
Hi Mathilde Borg Dahl, this is rather a general problem than a problem that is specific to SRS or rarefy. Reducing the number of read counts will always be accompanied with a loss of information.
I can't give you any recommendations because I don't know your study design as well as the type of organisms and environment that you sequenced. Feel free to contact me by email for further discussion regarding your data.
Cheers, Lukas
Thank you for your fast reply. I will discuss a bit with Haitao (we are in the same group), maybe I will write a follow up.
As of right now, I think I will do SRS for the richness evaluations and use within-sample relative abundance for the rest of my community evaluation. The dataset is bacteria and fungi (amplicon seq).
Sounds like a plan!
just a short note:
SRS is now on CRAN https://CRAN.R-project.org/package=SRS
Please use version 0.2.1 (not 0.2.0 as it contains a little bug that we removed).
Our package also features a Shiny app that you can use to explore different Cmin.
If you have any questions regarding SRS, feel free to contact me.
So, I just looked at the richness in my raw data, the SRS normalized and a rarefied version - the two latter normalized to the smallest sample size (1,500 reads).
I attach the overview here.
SRS changed very little, only the top richest samples have 'lost' one (max two) 'taxa'.
The rarefying had a greater impact on richness and more samples have lost one to max. seven taxa.
The richness in the 'low richness'-samples are unchanged for both normalization methods.
I am not convinced that one method is fare better than the other.... But maybe one should simply NOT report richness for this type of data (even though I would really like to), but instead look at some diversity indices - as Haitao Wang recommended me just now :)
You achieved nice results with both methods. It seems like your samples aren't that diverse too, thus, normalising to 1,500 read counts isn't such an issue in your case. In other words, 1,500 read counts/sample are apparently sufficient to detect almost all species in your samples. I agree with you regarding the reporting of species richness because in some cases, species richness is a function of the sequencing depth.
Much success with your analysis! :)
Haitao Wang I met this same problem that rarefaction throw many rare taxa, so I proposed an "Average Rarefied Table". You can use it in QIIME2, https://library.qiime2.org/plugins/q2-repeat-rarefy/33/
It's better to do rarefaction, check the following paper: Preprint To rarefy or not to rarefy: Enhancing microbial community an...
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size