DNA barcoding aquarium cyprinids

[racoll]

Open access (free download) article in PLoS ONE: Barcoding and Border Biosecurity: Identifying Cyprinid Fishes in the Aquarium Trade.

If anyone is interested, there is a large amount of supplementary information associated with this paper, including a huge table of morphological data on distinguishing aquarium cyprinids, and Web-based interactive trees where taxon names are clickable, and the links take you to database pages with photos of the fishes.

In addition, the images are all Creative Commons licenced, so can be copied and altered freely providing the source is acknowledged.

Three PC members get mentioned too!

[Suckermouth]

Congratulations on the publication!

[Mike_Noren]

Congratulations!

[racoll]

I am not a fan of NJ at all (or any tree-based method for ID for that matter), and only included it to reflect how poorly it performs (see below). I've actually got another paper in prep. criticising tree-based methods heavily. It's my opinion that BI or MP do nothing to resolve the problems with trees (see here and here).

Many barcoding studies employ terminology describing, for example, species forming “cohesive clusters” differentiated from one another by greater interspecific than intraspecific divergence, i.e., the barcode gap of Meyer and Paulay [82]. In our study, we measured clustering in terms of monophyly in NJ phylograms, a tree-based method which performed well on data generated here, but suffered when combined with GenBank/Bold information. This method requires strict monophyly of each species, resulting in a situation where the inclusion of a single misidentified specimen renders all queries in that species as misidentifications. Although alternative tree-based measures are available (e.g. Ross et al. [39]), the use of NJ trees in general is questionable due their method of construction [29], [37] and topological uncertainty [37], [89]. Furthermore, for a variety of reasons, “good species” may not always be monophyletic at mtDNA loci, so this method may fail to recognise species with either a history of introgression, or young species with large effective population sizes retaining ancestral polymorphisms [49], [73], [90]. These problems are not resolved through the use of bootstrap values, as we observed a significant reduction in identification success rate when node support was considered (up to 10%); recently divergent sister species on short branches were often not supported, even if they were monophyletic and diagnosable. DNA barcoding aims to maximise congruence between morphological identifications and sequence information while minimising misdiagnosis, but this is seriously undermined when bootstrap support values are included. For the reasons stated above, NJ trees are best avoided as a sole identification method [91], but can be a useful way to visualise and summarise patterns within barcode data.

[racoll]

Ah, you removed the comment. Please put it back. I enjoy the discussion!

[Silurus]

In the light of the recent Sundadanio review, which species do you think you used in your study?

Any reason why you use the taxonomy of Tang et al. (2010) over Liao et al. (2009)?

[Suckermouth]

In the light of the recent Sundadanio review, which species do you think you used in your study?

Man I didn't know about that. Can you send me that Sundadanio review? PM me if you forget my e-mail address. I'm collecting literature on danionines and I'm sure that'd be a necessary paper...

[racoll]

In the light of the recent Sundadanio review, which species do you think you used in your study?

I am not sure, yet. I'll be submitting my thesis in a matter of weeks, so it's not terribly high on priority list at the moment. But once things calm down, I'll get the specimens out, have a good read of the paper, and update the records online.

Any reason why you use the taxonomy of Tang et al. (2010) over Liao et al. (2009)?

Nothing personal to either of them, but I thought Tang et al.'s argument was stronger. I don't think molecular data has primacy over morphology, or vice versa, so taking a conservative approach and synonymizing them into Rasbora until further work is carried out is, to me at least, the best idea.

[racoll]

Man I didn't know about that. Can you send me that Sundadanio review? PM me if you forget my e-mail address. I'm collecting literature on danionines and I'm sure that'd be a necessary paper...

I compiled a huge amount of literature on danionines (including some very hard-to-get stuff), and it's all listed in the Table S1. If you need any of them, let me known, and I can send them to you.

[Silurus]

The Sundadanio review is available here.

[Suckermouth]

Man I didn't know about that. Can you send me that Sundadanio review? PM me if you forget my e-mail address. I'm collecting literature on danionines and I'm sure that'd be a necessary paper...

I compiled a huge amount of literature on danionines (including some very hard-to-get stuff), and it's all listed in the Table S1. If you need any of them, let me known, and I can send them to you.

Fantastic, I'll let you know if there's anything I want!

The Sundadanio review is available here.

An IEF paper online for free? Unthinkable!

[racoll]

An IEF paper online for free? Unthinkable!

It's quite sad that it's now easier with BHL to get hold of a paper from the 1800's than that journal.

How much longer they can continue with that kind of business model, I do not know. Frankly speaking, I think names should no longer be considered available unless they are published in open access journals. Biodiversity information is just too important.

ZooKeys are now doing very competitive rates, as are Zootaxa.

[Shovelnose]

Excellent!!!! Congratulations Rupert!!!

[MatsP]

Congratulations, Rupert!

I look forward to reading that later on.

How much longer they can continue with that kind of business model, I do not know. Frankly speaking, I think names should no longer be considered available unless they are published in open access journals. Biodiversity information is just too important.

Never mind the argument that nearly all of this sort of research is at least largely publicly funded, and the "result" is then to be paid for, and the profit of this is going to a private company.

--
Mats

[Mike_Noren]

I am not a fan of NJ at all (or any tree-based method for ID for that matter), and only included it to reflect how poorly it performs (see below).

Ah. I thought it was because that's what BOLD decided to use; BOLD in turn decided to use it partly because they appear to have been unaware of its drawbacks, but mainly because it's computationally simple and wouldn't bog down their servers.

I've actually got another paper in prep. criticising tree-based methods heavily. It's my opinion that BI or MP do nothing to resolve the problems with trees

Well, if you're trying to determine how similar A is to B, then doing a cluster analysis to check how similar A is to B to me would seem a reasonable approach - BLAST, for instance, is good at identifying barcodes (as is BOLD provided there aren't too many misidentified sequences). If you're trying to reconstruct phylogeny then a phylogeny reconstruction method would seem a better choice. My comment was related to Neighbor-Joining being a bad method of phylogenetic reconstruction - it's more sensitive to Long Branch Attraction than parsimony, more sensitive to model violations than Bayesian or ML, it's incapable of dealing with missing data, and the result is influenced by taxon input order unless you bootstrap with randomized taxon order (and then it loses its speed advantage). I bet you get a different tree if you use Bayesian or Maximum Likelihood analysis with the same model you used for NJ (that is typically the case except with very clean data). I also bet if you run ModelTest on your data it will not suggest K80 as model.

These problems are not resolved through the use of bootstrap values, as we observed a significant reduction in identification success rate when node support was considered (up to 10%); recently divergent sister species on short branches were often not supported, even if they were monophyletic and diagnosable.

That is expected, as the point of resampling is to detect problems in the dataset by collapsing nodes supported by conflicted data (in the specific case of NJ resampling also reduces input-order dependence). Resampling effectively makes your phylogenetic hypothesis more conservative, and while this minimizes the amount of ad hoc explanations in a phylogenetic paper, erring on the side of caution is not necessarily desirable when you want to find out if the shipment of pickled eel contains endangered Anguilla anguilla.

(Also a small nitpick: when determining monophyly from trees there is only monophyly and non-monophyly. The terms paraphyly and polyphyly are only applicable to character-based analysis; there is no qualitative difference between paraphyly and polyphyly in a tree.)

NJ trees are best avoided as a sole identification method [91], but can be a useful way to visualise and summarise patterns within barcode data.

Sure. I use NJ for quick-and-dirty estimates when working with sequence data because it's always available and the analysis nearly instant, but I do my analyses in TNT and MrBayes.

[Jools]

It is wonderful to read the culmination of hard work and this, it would appear, is quite a lot of it! Where can I find the interactive web based stuff? Sorry if I missed it.

Cheers,

Jools

[MatsP]

The link to the interactive one is in the link to the "supplementary material".

http://rupertcollins.net/trees/RHO.html

--
Mats

[racoll]

Where can I find the interactive web based stuff? Sorry if I missed it.

Yeah, it's all a bit dense I'm afraid. I was in two minds to break it up into more papers, but as they say, "a bird in the hand is worth two in the bush".

Thanks for posting the link Mats. That's to the smaller rhodopsin tree. The much bigger COI tree is here.

I'd appreciate some feedback on whether people are able to view/click the tree properly on their machines/browsers. I know it does not work properly on IE, but really there was no way round this. I had to either accept that, or not use the method. To open a taxon in a new tab, just hit ctrl-click. For some reason you can't do it with a right-click. The links with photos associated with them begin "RC0". The rest are from GenBank.

If anyone is interested in any of the cyprinid stuff, I posted a bit more detail at seriouslyfish.com. I thought they might be more into that, but I'm happy to discuss it here too.

Ah. I thought it was because that's what BOLD decided to use; BOLD in turn decided to use it partly because they appear to have been unaware of its drawbacks, but mainly because it's computationally simple and wouldn't bog down their servers.

We don't actually know what BOLD does, because it has never been documented properly. My interpretation of it, as I implemented locally for the paper, was to rank all closest matches (everything within 1% of your query), and return the ID if there was only one species present (otherwise as ambiguous for >1 species).

If you're trying to reconstruct phylogeny then a phylogeny reconstruction method would seem a better choice. My comment was related to Neighbor-Joining being a bad method of phylogenetic reconstruction

I agree, but as you say, identification does not hinge on phylogeny. Interestingly, it has been reported several times that for identification, NJ outperforms ML, MP, and BI. Still, there's not a lot in it, and they are all among the worst performing methods anyway. It's the best of a bad lot!

I also bet if you run ModelTest on your data it will not suggest K80 as model.

I bet that that if you run ModelTest on thousands of individual species, then not once will it select K2P/K80. F81 and HKY yes, but never K2P. Not that it matters though, as model selection has no impact on identification success. Raw p-distances will in fact perform marginally better than GTR etc.

[Shovelnose]

Did you manage to examine any specimens of P.exclamatio??? If yes, what do you think of the comments made in the P.rubrotinctus paper on this species.


"The hybrid Puntius clustered close to P. arulius in the COI NJ phylogram (Figure S1), while it clustered with P. denisonii in the RHO phylogram (Figure S2). This result indeed supports its identification as a hybrid, and potentially identifies the parental species."

Where where the hybrid specimens from??? I did not get this fully but does it speculate that P.denisonii could be hybridising in the wild??? Could you dumb this para down a bit for me please.

[Mike_Noren]

We don't actually know what BOLD does, because it has never been documented properly.

I have no insider information on that, but I attended a startup conference at Guelph, and the only method of analysis presented was neighbor-joining, with Hebert holding a presentation on its supposed advantages over other methods of analysis. Several people suggested other methods of analysis, but Hebert argued that other methods would be too taxing for the servers.

Interestingly, it has been reported several times that for identification, NJ outperforms ML, MP, and BI. Still, there's not a lot in it, and they are all among the worst performing methods anyway. It's the best of a bad lot!

No, it's the second-worst of the wrong set of tools. The only way NJ could outperform other model based methods at anything is if a) taxon input order accidentally results in the right tree, b) long branch attraction accidentally results in the right tree, or c) NJ uses a different model.

model selection has no impact on identification success. Raw p-distances will in fact perform marginally better than GTR etc.

Of course model selection has impact on identification success; you presumably mean that more complex models do not outperform simple models.
I do not doubt that, as phylogeny reconstruction has a different goal than barcoding: in phylogeny reconstruction you are interested in relationships over evolutionary time, in barcoding solely in degree of similarity among present diversity. As barcoding has no evolutionary pespective, applying corrections serves no purpose and the appropriate set of tools for barcoding is cluster analysis, not phylogeny reconstruction.

[racoll]

Did you manage to examine any specimens of P.exclamatio???

Sadly not, as the only ones I got from that group were P. arulius, P. tambraparnei, P. assimilis and P. filamentosus. There was a fair bit of diversity within P. assimilis, but the specimens were young, so not a lot more information could be found there.

Where where the hybrid specimens from??? I did not get this fully but does it speculate that P.denisonii could be hybridising in the wild???

They were hybrids bred deliberately for the trade, so not wild. They've been around for a few years, but luckily aren't too common. People assumed they were P. everetti x P. denisonii, but they are actually a hybrid between P. denisonii and something very close to P. arulius. The odd thing is you'd assume it was one of the common ones, but I collected all those, and it wasn't any of them.

Hybrid:

[racoll]

No, it's the second-worst of the wrong set of tools. The only way NJ could outperform other model based methods at anything is if a) taxon input order accidentally results in the right tree, b) long branch attraction accidentally results in the right tree, or c) NJ uses a different model.

Despite these well known problems with NJ, it has been documented empirically several times over, that it performs as well or slightly better than ML and MP (see here, here and here). I don't think anyone has tested BI, but I would expect it to be worse than ML, as it comprises a consensus of thousands of trees, which will immediately reduce resolution. Given the massively faster analytical times of NJ, I don't know why anyone would use MP, ML or BI for this purpose.

As we've both acknowledged though, tree-based methods should not be used anyway. Little (2010) summed it up nicely, saying methods "that do not consistently perform as well as pairwise matching are manifestly flawed. Thus, the data presented here suggest that due to inconstant performance no tree–based method should be used for barcode sequence identification.".

Several people suggested other methods of analysis, but Hebert argued that other methods would be too taxing for the servers.

That's utter nonsense, as I'm sure you're aware. A far simpler method like BCM is faster than even NJ, offers significantly (in the statistical sense) better results, and has more desirable properties. Even BLAST is better than NJ, but I don't like that much either, as the outputs can be quite ambiguous.

Of course model selection has impact on identification success

It really doesn't. There is almost zero difference between models in terms of ID success. The most complex models (i.e. with gamma parameters) are very, very slightly worse than the rest, but there is basically nothing in it. The zone in which identifications are made (i.e., < 2%) is too small to be corrected by any of the standard models.

[Shovelnose]

They were hybrids bred deliberately for the trade, so not wild. They've been around for a few years, but luckily aren't too common. People assumed they were P. everetti x P. denisonii, but they are actually a hybrid between P. denisonii and something very close to P. arulius. The odd thing is you'd assume it was one of the common ones, but I collected all those, and it wasn't any of them.

Thanks. Yeah, it is strange P.arulius is in the mix considering they are hardly seen in the trade.

I thought hybrid (P.filamentosus * P.assimilis especially) specimens were common in the trade. Did you come across any wild caught hybrids??? I have seen posts by hobbyists from England on another forum and there were some really weird looking wild caught Puntius available a few years ago.



Puntius assimilis and P. tambraparniei spawned naturally at a friends place last year.

[racoll]

Did you come across any wild caught hybrids???

No, none that I noticed, but I dare say it happens frequently enough.

Quite an amorous looking P. tambraparniei there!

[Shovelnose]

Quite an amorous looking P. tambraparniei there!

Ha ha. True.



Pic Credit : Mick Wright

One of the weird ones I was talking about.

[Mike_Noren]

Despite these well known problems with NJ, it has been documented empirically several times over, that it performs as well or slightly better than ML and MP

For barcoding, a task which ML and MP (and NJ) simply are not designed to do. It is, like you said, a case of being the least-worst at a job it was never designed for, like being the least-worst type of saw for hammering nails.
That said, empirically my experience is that with real data NJ is more likely to misplace even conspecific terminals than at least Bayesian analysis is.

As we've both acknowledged though, tree-based methods should not be used anyway.

"Tree based" is not a good term, but yes, methods designed for reconstructing phylogeny are not a good choice for identification through barcoding.

That's utter nonsense, as I'm sure you're aware.

Yes. To be fair the statement was made in the context of other methods of phylogenetic analysis such as parsimony.

A far simpler method like BCM is faster than even NJ, offers significantly (in the statistical sense) better results, and has more desirable properties. Even BLAST is better than NJ, but I don't like that much either, as the outputs can be quite ambiguous.

Really? I think it performs quite well, what is it you don't like about it?

The zone in which identifications are made (i.e., < 2%) is too small to be corrected by any of the standard models.

It was a semantic nitpick, and I apologize: it is possible to design models which give ridiculous results, hence choice of model is not irrelevant.

[racoll]

Really? I think it performs quite well, what is it you don't like about it?

Mostly the ambiguity really. If the taxon one is querying does not have a conspecific reference in the database, or is divergent to some degree from the reference, then interpreting the result can be difficult. The percent values given are not the same as a standard genetic distance, so threshold values may not be comparable. There are also lots of settings which may influence the result somewhat. It's another tool that isn't really designed to do the job of barcode IDs, but it usually performs well, and its a darn sight better than NJ ...