Hello,
I understand there's disagreement regarding the similarity threshold for fungal OTUs, be it 95%, 97%, or 90-99%. In the literature, anyone of these threshold may be cast as adequate or too conservative.
For my thesis, I want to determine the fungal communities among tree-samples with next-gen sequencing.
From your own experiences, which similarity threshold(s) were most appropriate?
Thank you,
Gary
Edit: To clarify, I am using the ion torrent platform Ion S5 for sequences 200-400 bp long. Compared the Ion PGMʻs 5 million sequencing capacity, the S5 530-chipʻs maxes out at 20 million sequences.
I am very much on board with using sequence variants (i.e. ESVs, ASVs, ZOTUs are all terms used) as opposed to clustering sequences by similarity. My rationale is based on both biology and technical concerns.
1. As you mentioned, the similarity threshold one chooses is largely arbitrary and determined via consensus among practitioners. In some cases, taxa that are >97% similar ARE functionally different and have been labeled different things by taxonomists, and in other cases the 97% threshold works ok for delineating "species". There is not a whole lot of guidance out there for when the threshold works and when it doesn't. If you are working in a remote system and sequencing things nobody has sequenced before, then your guess is as good as anyone's regarding how well the 97% threshold captures variation at some specific level of the biological hierarchy. To sum up, the arbitrary nature of the 97% threshold is unsatisfying.
2. Depending upon the length of your amplicon, a percentage based threshold will require more or fewer bases to generate a new OTU. E.g. for a 100 bp long amplicon 4 bases need to differ between two sequences for them to be >3% different, but for a 200bp long amplicon 7 bases do. The denominator changes in the percentage calculation. For fungi in particular I think this is an issue because the ITS region is variable in length, so the 97% threshold (or any other threshold) likely means different things across the fungal phylogeny.
3. If you use a similarity threshold then you can inflate richness estimates and community membership slightly within a sample. Say for instance you have two 100bp sequences in a sample (call it sample 1) that are identical except one has AAAA and the other has TTTA at the same locus. If you processed this sample by itself these two sequences would probably get assigned to the same OTU because they are 97% similar. However, further suppose that you are processing several samples together and in one of these other samples you call an OTU that has the sequence TTTT at our focal locus. Now, if we processed sample 1 with this extra OTU sequence in our consensus OTU database we would find that two OTUs are in sample one! To sum up, the similarity threshold can lead to inflated OTU estimates in some cases when processing more samples together (which we all do) than if you processed those samples independently.
4. When using similarity thresholds one throws away biologically meaningful variation. One can use relative abundance of exact sequence variants to learn about genetic variation within a taxon. In my mind, the goal is to learn about biological variation, what led to that variation, and what are the consequences of that variation to communities. Therefore we should retain all our information. This is the Bayesian in me....
You could check out a paper I just published for an example of analyzing ESVs within a focal taxon to link genetic variation to differences in the production of a focal metabolite (Harrison et al. 2018, J of Ecology). I found the ESV approach very rewarding here as it let me learn something regarding the consequences of genetic variation within a fungal taxon that I would have missed if I had only used the standard 97% similarity threshold
5. Using exact sequence variants better facilitates cobbling data together from disparate labs, methods, etc. because one has the exact sequences instead of a nebulous OTU that hides biologically meaningful variation. For a cool example of how ESVs can be useful see:
Thompson, L. R., Sanders, J. G., McDonald, D., Amir, A., Ladau, J., Locey, K. J., ... & Navas-Molina, J. A. (2017). A communal catalogue reveals Earth’s multiscale microbial diversity. Nature, 551(7681).
6. A criticism of ESVs is that if you look at the genome with enough resolution everything becomes different. However, we must remember that by using standard primer pairs we limit the portion of the genome under consideration. Therefore, I do not think this criticism holds water for amplicon based work.
For all these reasons I do not think it is relevant for us to continue using similarity thresholds for OTU delineation.
See Callahan et al. for more thoughts on this: Callahan, B. J., McMurdie, P. J., & Holmes, S. P. (2017). Exact sequence variants should replace operational taxonomic units in marker-gene data analysis. The ISME journal, 11(12), 2639.
I am keen to hear counterarguments from folks!
I am very much on board with using sequence variants (i.e. ESVs, ASVs, ZOTUs are all terms used) as opposed to clustering sequences by similarity. My rationale is based on both biology and technical concerns.
1. As you mentioned, the similarity threshold one chooses is largely arbitrary and determined via consensus among practitioners. In some cases, taxa that are >97% similar ARE functionally different and have been labeled different things by taxonomists, and in other cases the 97% threshold works ok for delineating "species". There is not a whole lot of guidance out there for when the threshold works and when it doesn't. If you are working in a remote system and sequencing things nobody has sequenced before, then your guess is as good as anyone's regarding how well the 97% threshold captures variation at some specific level of the biological hierarchy. To sum up, the arbitrary nature of the 97% threshold is unsatisfying.
2. Depending upon the length of your amplicon, a percentage based threshold will require more or fewer bases to generate a new OTU. E.g. for a 100 bp long amplicon 4 bases need to differ between two sequences for them to be >3% different, but for a 200bp long amplicon 7 bases do. The denominator changes in the percentage calculation. For fungi in particular I think this is an issue because the ITS region is variable in length, so the 97% threshold (or any other threshold) likely means different things across the fungal phylogeny.
3. If you use a similarity threshold then you can inflate richness estimates and community membership slightly within a sample. Say for instance you have two 100bp sequences in a sample (call it sample 1) that are identical except one has AAAA and the other has TTTA at the same locus. If you processed this sample by itself these two sequences would probably get assigned to the same OTU because they are 97% similar. However, further suppose that you are processing several samples together and in one of these other samples you call an OTU that has the sequence TTTT at our focal locus. Now, if we processed sample 1 with this extra OTU sequence in our consensus OTU database we would find that two OTUs are in sample one! To sum up, the similarity threshold can lead to inflated OTU estimates in some cases when processing more samples together (which we all do) than if you processed those samples independently.
4. When using similarity thresholds one throws away biologically meaningful variation. One can use relative abundance of exact sequence variants to learn about genetic variation within a taxon. In my mind, the goal is to learn about biological variation, what led to that variation, and what are the consequences of that variation to communities. Therefore we should retain all our information. This is the Bayesian in me....
You could check out a paper I just published for an example of analyzing ESVs within a focal taxon to link genetic variation to differences in the production of a focal metabolite (Harrison et al. 2018, J of Ecology). I found the ESV approach very rewarding here as it let me learn something regarding the consequences of genetic variation within a fungal taxon that I would have missed if I had only used the standard 97% similarity threshold
5. Using exact sequence variants better facilitates cobbling data together from disparate labs, methods, etc. because one has the exact sequences instead of a nebulous OTU that hides biologically meaningful variation. For a cool example of how ESVs can be useful see:
Thompson, L. R., Sanders, J. G., McDonald, D., Amir, A., Ladau, J., Locey, K. J., ... & Navas-Molina, J. A. (2017). A communal catalogue reveals Earth’s multiscale microbial diversity. Nature, 551(7681).
6. A criticism of ESVs is that if you look at the genome with enough resolution everything becomes different. However, we must remember that by using standard primer pairs we limit the portion of the genome under consideration. Therefore, I do not think this criticism holds water for amplicon based work.
For all these reasons I do not think it is relevant for us to continue using similarity thresholds for OTU delineation.
See Callahan et al. for more thoughts on this: Callahan, B. J., McMurdie, P. J., & Holmes, S. P. (2017). Exact sequence variants should replace operational taxonomic units in marker-gene data analysis. The ISME journal, 11(12), 2639.
I am keen to hear counterarguments from folks!
Amplicon (e.g. ITS) sequencing can never be guaranteed to give one the absolute "truth" in terms of species abundance (even if by variant analysis e.g. dada2) , but if you apply the same criteria across all your samples the relative answer will be correct. (e.g. group A has higher species richness than group B, the taxa in A overlap to what extent with group B etc) independent of sequence variant or OTU based approach. Sequence variant approaches infer particular types of PCR errors, but this can be impacted by the particulars of the wet bench chemistry used. Sequencing platform (and associated error types) are also an issue. If you can do a sequence error correction for overlap from forward and reverse reads it will help as well (you can get overlap with Miseq 2X 300 reads on most fungi for ITS 1 will give you complete sequence in both directions and allow for sequence error correction (e.g Mothur Illumina MiSeq SOP https://www.mothur.org/wiki/MiSeq_SOP ).
I agree in principle with the previous 2 answers, and will briefly add that the accuracy depends on the combination of amplicon length, gene of interest, reliability of reference accessions AND % similarity. If you are comparing a group of relatively well-known fungi then a higher % threshold for taxa determination will obviously be more appropriate, but less well known and novel organisms require a broader assessment.
- Badges
- Science method