Discrepancies in Tree Topology with Different Models in IQ-TREE

[Adrià Diéguez Moscardó]

Hi everyone,

 I'm encountering an issue with tree topology discrepancies when using different models in IQ-TREE and I'm looking for some suggestions.

Considering a quartet with branch lengths 0.1, for example, (Leaf_1:0.1,Leaf_2:0.1,(Leaf_3:0.1,Leaf_4:0.1):0.1),  I've used an alignment simulator to generate alignments given this topology and these branch lengths.

Then, I have provided the simulated FASTA file and the tree topology as inputs to IQ-TREE to estimate the branch lengths of the tree. When I use the GTR model (012345), I obtain very accurate results for branch length estimation, and the correct topology is returned. For example: (Leaf_1:0.099,Leaf_2:0.101,(Leaf_3:0.101,Leaf_4:0.096):0.102);

 However, using the 12.12 model, I notice that IQ-TREE returns different topologies, such as:  ((Leaf_1:0.100,Leaf_2:0.094):0.055,(Leaf_3:0.102,Leaf_4:0.100):0.049):0.000 or ((Leaf_1:0.095,(Leaf_3:0.098,Leaf_4:0.099):0.104):0.088,Leaf_2:0.000):0.000;

 I also noticed that the output with the GTR model (012345) gives an unrooted tree, whereas the 12.12 model always returns a rooted tree and places a root. Could this be contributing to the observed discrepancies?

Any help or guidance would be greatly appreciated.  I'm particularly interested in understanding why the 12.12 model leads to different topologies. Thank you in advance!

Best regards,

Adrià

[Sishuo Wang]

Dear  Adrià,

I am not an expert on it. As follows i try to share w you some of my perhaps naive thoughts.

You are right that 12.12 always returns a rooted tree. That's bc of its time non-reversibility, so unlike traditional reversible markov models, it always returns a rooted tree, which means diff root positions, while keeping any other parts of the tree unchanged, could (probably) result in diff likelihoods . If i am not mistaken, people often use AIC or BIC or AICc as implemented in iqtree for empirical data sets where u don't know the true phylogeny. To calculate it, you might want to consider not only the log-likelihoods, but the # of params in the model. I think 12.12 has 11 free model parameters to estimate, while GTR has 8 (bc it has to ensure q_ij * π_i = q_ji * π_j); in addition, 12.12 has one more for the branch lengths (as it has a root) than GTR. But actually u can directly find all of the AIC etc. info in ".iqtree" file for model selection. You might also want to note the warning if you have too few nucleotides (n) as observed in the file .iqtree.

In your case you are using simulations and that could give your more clues to which model is better to use. More info, you might wish to visit http://www.iqtree.org/doc/Substitution-Models#lie-markov-models, http://www.iqtree.org/doc/Rootstrap, https://academic.oup.com/sysbio/article/71/4/959/6350503

best,

sishuo