Low quality phylogeny

[Giuseppe Emanuele De Pasquale]

Hello everyone,

I am encountering some issues with a mitochondrial phylogeny of canids (primarily wolves and dogs) due to low bootstrap support. I conducted an alignment of approximately 800 sequences encompassing all 13 mitochondrial genes, rRNA, tRNA, and D-loop regions concatenated. The alignment underwent trimming and filtering, wherein intergenic sites were deleted, and sequences with more than 15% gaps were removed. Subsequently, I generated haplotypes using DNAsp, including gaps (to potentially account for haplotypes with indels) and invariant sites. The final alignment consists of 543 sequences with a maximum length of 16433 bp.

I attempted to create six different phylogenies by segmenting the alignment as follows:

GENES
GENES + D-loop
GENES + rRNA
GENES + rRNA + D-loop
GENES + rRNA + tRNA
GENES + rRNA + tRNA + D-loop

For each segmentation, I employed a partition scheme generated by IQ-Tree 2.2.2.7 on CIPRES, testing each codon position for each gene and separating non-coding regions. Here is an example from the latest phylogeny:

charpartition mymodels =
  TN+F+I+R2: ATP6_pos1_ATP8_pos1_ATP8_pos2_COX1_pos1_COX2_pos1_COX3_pos1_Cytb_pos1_ND1_pos1_ND2_pos1_ND3_pos1_ND4L_pos1_ND4_pos1_ND5_pos1_12S_16S_22_tRNAs,
  TN+F+I+R2: ATP6_pos2_COX1_pos2_COX2_pos2_COX3_pos2_Cytb_pos2_ND1_pos2_ND2_pos2_ND3_pos2_ND4L_pos2_ND4_pos2_ND5_pos2_ND6_pos2,
  TN+F+I+R2: ATP6_pos3_ATP8_pos3_COX1_pos3_COX2_pos3_COX3_pos3_Cytb_pos3_ND1_pos3_ND2_pos3_ND3_pos3_ND4L_pos3_ND4_pos3_ND5_pos3,
  TN+F+G4: ND6_pos1_ND6_pos3,
  TPM3u+F+R3: DLOOP;

The command used for each phylogeny was:

iqtree2 -nt 6 -bnni -bcor 0.99 -nm 1000 -s alignment.fasta -p partition.nex -st DNA -m TESTNEWMERGE --sprrad 6 -B 1000

I experimented with additional phylogenies using different partition schemes and IQ-Tree versions. However, I noticed that newer versions of IQ-Tree tend to provide similar model substitutions for every partition. In the latest phylogeny, I observed three identical models for three different partitions.

Despite these efforts, I consistently obtained similar bootstrap values across all phylogenies. The best is the last with support values from 40 to 60 on the principal nodes. Basal nodes have high bootstrap in every phylogeny.

Additional information:

-I attempted phylogenies without collapsing haplotypes, resulting in slightly worse outcomes.
-Using alignments with less than 5% gap sequences yielded similar results.
-The topology remains consistent across all phylogenies and aligns with other studies.
-I increased bootstrap iterations up to 2500.
-I cross-checked gene alignments using a reference of translated proteins to ensure accuracy in the partition file's codon coordinates.

My questions are:

-Why do older versions of IQ-Tree yield more heterogeneous results in model selection?
-Why do recent versions fail to merge identical models, and why are there identical models for different regions (invariant and variable sites)?
-How can I improve my bootstrap values? Could I be overlooking something?
-Is there a better version of IQ-Tree for this dataset?

Sorry for the lengthy question.
Any help, opinion, or experience is greatly appreciated.

Best regards!

[Heiko Schmidt]

Dear Giuseppe,

There are others who are certainly better in answering most of your questions here.
I just want to make a comment to two of your questions:

> -Why do older versions of IQ-Tree yield more heterogeneous results in model selection?

What older versions are you referring to?
I know that between IQ-TREE 1 and 2 there was a change in model selection to speed up model selection in large data sets. Models that cannot be expected to get higher scores than already tested ones, not all rate heterogeneity types are tested anymore.

This behaviour can be overruled by setting “-mrate ALL”.
Maybe that reduced some heterogeneity, although I don’t think so.

Minh and others might know more.

> -Why do recent versions fail to merge identical models, and why are there identical models for different regions

Please note, just because two sets of genes/regions have their same type of model (here TN+F+I+R2), it does not mean that their estimated parameterizations are substantially different.
Different codon positions or genes could substantially different GC contents or the relative rates in the substitution matrix can be distributed very differently. And in a +R2 the relativ rates of the two site categories could also vary strongly in the difference and size.

Hence, two TN+F+I+R2 models can actually be quite different, and then they should not be joined.

I hope that helps at least for one of your questions.

Best wishes,
Heiko Schmidt

> --
> You received this message because you are subscribed to the Google Groups "IQ-TREE" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to iqtree+un...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/iqtree/f00e6293-6f0e-4ec2-be37-984a490fd39fn%40googlegroups.com.

-----------------------------------------------------------------------------
Heiko Schmidt
Center for Integrative Bioinformatics Vienna (CIBIV)
University of Vienna / Max Perutz Labs
http://www.cibiv.at/
-----------------------------------------------------------------------------

[Giuseppe Emanuele De Pasquale]

Actually, I have found more heterogeneous patterns with version 1 of IQ-Tree, so your explanation seems plausible to me. Thank you for the clarification.

It seems to me, therefore, that the procedure I have followed should be fine, so I cannot understand why I am not getting good support despite having long sequences.

I await any further feedback.

Best regards!
Giuseppe

[rob.l...@gmail.com]

Hi Guiseppe,

Just a quick one - there is no sense in which some bootstraps are 'good' and others 'bad'. They are an estimate of the sampling variance in your data (well, 100% minus the sampling variance). If your bootstraps supports are low, this just suggests that the sampling variance in your data is high. In other words, when you resample the dataset using a bootstrap, you frequently get quite different trees.

The usual cause of low bootstraps is that there is a relatively small amount of information in your alignment. To get some idea of this you could count the number of informative sites (IQ-TREE outputs this). Much rarer (especially with mtgenomes) is that there's lots of information but it's in almost perfect conflict.

I hope this helps,

Rob

[Giuseppe Emanuele De Pasquale]

Hi Rob,

Thanks for the clarification!

Giuseppe

Privo di virus.www.avast.com

You received this message because you are subscribed to a topic in the Google Groups "IQ-TREE" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/iqtree/sxOU-Cm4NQM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to iqtree+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/iqtree/4cc57449-5ea0-4eef-a567-2d01c721d9f3n%40googlegroups.com.

[Bui Quang Minh]

Hi Giuseppe, Heiko and Rob gave very nice feedback. I'd like to echo
on Rob's point about phylogenetic information: you have long
sequences, but if many of them are very similar (e.g. having many
invariant sites in the alignment), then you may have low information.
IQ-TREE printed some statistics which may hint into this, such as
number of invariable sites, [parsimony] informative sites. In such
case: I would actually abandon (complex) partition models, because it
may overfit the data, and just use a single model such as GTR+I+G for
all sites in the alignment.

Another thing to consider: You can also try some rogue taxa
identification tool (taxa which are randomly moving around the tree
e.g. due to a long branch leading to it or it doesn't fit the
alignment). Removing these rogue taxa may improve bootstrap supports.
I believe one of the tools is called RogueNaRok from A. Stamatakis
group.

Cheers,
Minh

> To view this discussion on the web visit https://groups.google.com/d/msgid/iqtree/4cc57449-5ea0-4eef-a567-2d01c721d9f3n%40googlegroups.com.

[Giuseppe Emanuele De Pasquale]

Thanks, Minh!

I've identified around 50 rogue taxa using RogueNaRok, and I'm now planning to redo the analyses, excluding these taxa.

Hopefully, the results will be improved! Fingers crossed!

Cheers,
Giuseppe