Model Questions and Missing Information on Tutorial

[danielagandrade]

Hi CAFE5 developers, I am new to using your software and I am excited by the questions it is allowing me to ask. Thank you for developing!

I have a couple of clarifying questions that I would appreciate getting your input on. Additionally, I wanted to bring a couple of things to your attention.

1. Does CAFE5 work well with gene families that have several species with 0 gene counts or is it best to limit CAFE5 to orthogroups shared by all species? With Orthofinder, I have identified a total of 33,123 Orthogroups but only ~4300 are shared by all my species. These are the gene families I have chosen to include as CAFE's input, but I was wondering if I am loosing a lot of enlightening data by avoiding the appearance of zeros and therefore would like to know if CAFE is robust with 0 gene counts in gene families.

2. On CAFE5's manuscript (Mendés et al. 2021) there are a couple of important observations made regarding K that I wanted to clarify. It is said that K = 1 usually underestimates lambda and that K= 2-3 slightly over-estimates lambda. Is there a good way then of knowing how to select the number for K? I have seen on a couple of discussion posts that there is no clear way to determine which K is best and that the likelihoods of models with differing Ks cannot be compared. However, Fig.1A on the manuscript shows an * on the K number for which the highest likelihood was estimated. Therefore, I am a bit confused. Is the maximum likelihood a good indicator of a proper K? Or how does one determine that an optimal or "good enough" K has been found for the model?

3. Can models with varying lambdas be considered nested? In other words, can I do the likelihood ratio of models with 1-4 lambdas to determine what fits my data best? The species I am comparing are all from the same Class but some diverged over 200 million years ago, so I do not expect the rates to be the same across the entire phylogenetic tree.

4. Would you suggest determining the best K number first or the best lambda number first?

5. It is unclear to me how to use the Simulation function of CAFE5 and how it can help me determine which values are best for my model. When I look at the tutorial (http://github.com/hahnlab/CAFE5/blob/master/docs/tutorial/tutorial.md#simulation) , it refers to an example and to commands that are not shown:

"Simulation
Here, the genfamily command simulates the datasets (in the example above, we are asking for 100 simulations with -t 100). It estimates λ from the observed data to simulate gene families. Then the likelihoods of the two competing models are calculated with the lhtest function, which takes the multi-λ tree structure, and the estimated λ value using the global-λ model."

Is there another document I should refer to or do you have any suggestions?

Once again thank you for developing and for all of your help!

[hahnlab-user]

Hello,

Thanks for all your questions, and sorry for the lack of clarity in the tutorials. I will try to answer them here using your numbering scheme:

1. Yes, CAFE can work with gene families that have 0's in several species, with no issue. The only issue that can arise is with gene families that evolve de novo within a tree, such that the root state is 0--this is not modelled. As default, CAFE will throw out such families.

2. Unfortunately, there is no way to use the likelihood to decide on k. While the figure you refer to does show the k-value with the highest likelihood, the point of this is meant to show that this is not always the correct lambda-value.

3. Can models with different numbers of lambdas are usually nested, though this depends on how you assign lambda-parameters among branches (these assignments must themselves be nested).

4. I'm not sure there is a correct answer to this, mostly because CAFE currently cannot fit both gamma models and varying-lambda models simultaneously. (i.e. If you want to estimate a gamma model, you can only do this with a global lambda.)

5. Sorry, that part of the tutorial points to deprecated code. The simulation function is needed to generate the null distribution of likelihood values used to test nested lambda models. For updated commands, see here: https://github.com/hahnlab/CAFE5/blob/master/examples/all_cafe_commands.sh

cheers,
Matt

[danielagandrade]

Hi Matt,

Thank you for your clarifying answers and for the prompt response. I really appreciate it!