How to normalize bacterial abundance for unequal sample size?

Dear All,
Ive executed QIIME2 analysis for my four different group. Each group consists of unequal sample size.
Group A: 11
Group B: 9
Group C: 31
Group D: 12
and the sequence counts of each sample varies from 182,014 to 4,194.
And I got more abundance in Group C for most of the taxa.
In which way I can normalize the samples and get the abundance of taxa?
Should I do any FPKM/RPKM or log2 normalization method ?
Is there any R or some other tools to normalize OTU abundance.

Hi @steffi,

Which analyses did you perform on what level?

Second, is there a reason to believe what you’re seeing is untrue and that group C isn’t dominated by single OTUs? If you run a PCoA based on weighted metrics, does it make sense that group C clusters more tightly (shares more weighted metrics)? You could see this in boxplots or in a PCoA, depending on other factors in the data set.

Third, if you randomly select a subset of Group C, does the pattern hold? If 10 samples from Group C still show the taxa is dominant, that might suggest an actual effect, adjusted for sample size, rather than a group size effect.

I also want to add some background on normalization methods in microbiome, because its complicated (what isn’t) and it’s been a big piece of discussion the last couple of years… Waste not, want not: why rarefying microbiome data is inadmissible (PMID: 24699258) was one of the first papers to address issues with rarefaction for taxonomy-based analyses, as well as propose one of the first widely accepted models. However, importance of rarefaction in diversity metrics was reinforced in Normalization and microbial differential abundance strategies depend upon data characteristics (PMID: 28253908), which is also worth a read. Finally, I recommend Microbiome Datasets Are Compositional: And This Is Not Optional (doi: https://doi.org/10.3389/fmicb.2017.02224). I think QIIME2 has tried to address the issue around compositionality in the tests it uses.

Best,
Justine

5 Likes

Hello @steffi and Justine,

We cannot compare relative abundance of a taxon between groups; ANCOM paper explains the reason (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4450248/). I wonder if comparing ratio of pairwise components between groups, which is calculated by ANCOM, is the most appropriate so far. Precisely, expected abundance of a taxon appears to be the best…

Bests,
ohmiyajohn

Hi @ohmiyajohn,

I think probably the best would be an internal standard. Noah Fierer’s group at UC Boulder used this approach in a recent paper, as well as an extra cellular DNA removal method(PMID: 27991881). And, looking for that article, i found another I really need to read: Absolute quantitation of microbiota abundance in environmental samples (PMID 29921326). The problem with these methods is that they require normalisation before sequencing, and most of us are dealing with post sequencing datasets. Ive been trying to convince my lab to do the spike in, but I think until Im doing my own extractions and sequencing, I may be limited in efficacy. These techniques essentially “break” compositionality because then everything can be put as a ratio against a constant, rather than a ratio against each other.

Best,
Justine

2 Likes

Hello Justine,

Thank you for your reply.
Although I looked through these papers you recommended, I could not find the evidence how accurate these methods measure the abundance …? Also, I am wondering if the “spike” method could be conducted for gut microbiota.

Since it is almost impossible to quantify the absolute abundance of taxa from samples we grab, the matter is how to predict the pseudo abundance of taxa in the ecosystem from our samples. Why I suggest the ratio of pairwise components calculated by ANCOM is because I use ANCOM to test the significance of differences in the abundance between groups. The box plot based on the ratio does not contradict the result of the significance test.

Best regards,
ohmiyajohn

Hi Justin,
Thank you for the reply. I performed analysis upto taxonomic classification using GG database(without alpha and beta diversity)

Yeah. In group C, the total number of sample is high (n=31) compared to other groups (11,12,9). Obviously the total abundance is more. Ive performed PCA analysis. In that group C did not cluster together. part of them clusters separately.

No. not all the samples contribute same abundance. We do not want to eliminate any sample. Thats y we are looking for normalization method. Is there anyway we can get normalized OTU table?

Hi @ohmiyajohn
Thank you for the reply. I have read the tutorial in QIIME2 which says that ANCOM identifies differentially abundant species across sample groups. Will that calculate each and every taxa present in the groups>?

Hi @ohmiyajohn,

So, this is not absloutely abundance, and that I should have specified. For absloute abundance, you’d need qPCR, which a lot of people run. Whole community PCR is difficult, but possible. Ive been out of the wetlab for a few years, but it should in theory be possible.

The idea of the spike-in is to break compositionality by providing something that represents a constant across all samples. So, rather than asking about the total ratio (like ANCOM) or the ILR-transformed balances, like in Gneiss and PhILR, you’re asking about whether or not each member of the community changes with regard to a constant. When you divide out the constant, you’re left with data that no longer conforms to the compositional limitations.

Since its a general principle for communities, it should work well. I’d pick a spike-in that isn’t present in my community. So, a synthetic 16s rRNA sequence or a marine organism that isn’t typically present but growable in the lab. That way, you know which is your reference easily.

Cheers,
Justine

1 Like

Perhaps I misunderstood the original question, then. I understood that the total number of samples is higher, and that these samples are dominated by a specific set of organisms which are at higher relative abundance than in the other groups. And that you were wanting to normalize for sample number across groups.

If you have a relative abundance table, you’ve already normalized the data. There are a lot of normalization approaches (as referenced in the papers I sent you), however, as @ohmiyajohn said and the review I linked says, your microbiome data is compositional and should be analysed that way.

ANCOM will tell you which of the features (ASVs, OTUs, genera, etc) are different between the two groups based on the compositional data.

You determine the taxonomic classification of each sample with the taxonomic classifer, whether you present the data as raw counts, rarefied data, a relative abundance table, or a compositional table.

Best,
Justine

1 Like

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