[Feature Request] Feasibility of implementing standard non-parametric bootstrap for CMAPLE / MAPLE

[Janzulene]

Background

MAPLE, as a fast phylogenetic inference method for large-scale data, has been integrated into IQ‑TREE 3. However, the standard -b bootstrap option does not work when MAPLE is used as the inference method((the same applies to the -o option for specifying an outgroup). At the very least, IQ‑TREE should at least issue a warning to users indicating that the bootstrap option will be ignored in this combination.

Furthermore, CMAPLE currently provides SPRTA as a branch support metric. In my own preliminary experiments, SPRTA scores show a low correlation with the branch support values obtained from the standard bootstrap ( -b / -bb ) under the traditional maximum‑likelihood (ML) framework in IQ‑TREE while the MAPLE tree itself shows a stronger topological correlation with the ML tree than other fast methods.

Suggestion

I would like to propose enabling Felsenstein’s non‑parametric bootstrap for CMAPLE. The basic workflow would be:

1. Perform column‑wise resampling (with replacement) of the input alignment.

2. Convert each bootstrap replicate into the MAPLE‑compatible format (e.g., using createMapleFile.py).

3. Run CMAPLE independently on each replicate.

4. Map the resulting topologies back to the original tree to compute branch support.

MAPLE is already substantially faster than traditional ML methods. Adding a bootstrap procedure would, of course, increase total computation time, but for large datasets it would still be much faster than running ML + bootstrap. If future experiments confirm that bootstrap is indeed more reliable than SPRTA (as my small‑scale tests already hint), then implementing bootstrap for MAPLE would be the best available option until an even faster method that can achieve near‑bootstrap reliability appears.

Adding standard bootstrap would also align MAPLE with other mainstream phylogenetic methods (ML, MP, etc.) in terms of branch support assessment, making it easier for users to compare and trust the results.

Major Obstacle & Open Question

Currently CMAPLE’s initialization logic expects the “reference + SNP difference” format. It remains unclear whether the algorithm is statistically compatible with resampled (bootstrap) data in terms of bootstrap consistency. Has any theoretical or simulation‑based validation been conducted?

Supporting Experiment

I have performed a small‑scale test: on a modest test tree, I compared SPRTA values against bootstrap / ultrafast bootstrap support for the same clades (identical topology). The correlation was not very strong. I plan to extend these experiments to larger simulated and real datasets and will share the detailed results in a follow‑up discussion. (Please note that due to other commitments, I may not be able to provide the full experimental follow‑up very soon — I appreciate your patience.)

[bqminh]

@trongnhanuit @NicolaDM what do you think?

[NicolaDM]

Hi @Janzulene ,
Thanks for the interest and for trying out the methods.
I wanted to point out 2 things about the Felsenstein bootstrap (FB) before discussing the proposed feature.

1. the FB does not measure the same thing as other branch support measures (like posterior clade support in Bayesian phylogenetics, SPRTA, or transfer bootstrap). This is not a defect of the other branch support measures, they are simply measuring different things, as they are meant to do. The meaning of the SPRTA scores is explained in the SPRTA paper.

2. the FB support scores have a number of limitations, and I don't think they should be considered ideals that other branch support scores should aim for. In addition to the SPRTA paper, this is also discussed for example in the TBE paper and the Bayesian bootstrap paper.

Personally I think that, in the context of short-divergence phylogenetics (which is the case use of CMAPLE), methods like SPRTA, TBE, the Bayesian bootstrap, and aBayes make more sense than the FB, due to the fact that the FB gives unintuitively low support to very short branches (because it measures repeatability and not posterior probability), due to the fact that it's affected by rogue taxa, and due to its very high computational demand.
CMAPLE already includes the SH-aLRT which should give results much more similar to the FB than SPRTA; additionally, unlike the FB, the SH-aLRT should be fast and robust to rogue taxa. @Janzulene if you are looking for a branch support measure more similar to the FB already in CMAPLE, I would suggest to consider the SH-aLRT.
If another branch support measure similar to the FB should be implemented in CMAPLE, I personally would lean towards aBayes (which like the SH-aLRT should be fast and robust to rogue taxa, but additionally shouldn't give excessively low support to short branches).

Of course these are just my opinions, and I'm sure other people in the field would argue differently.
Please feel free to argue here against these opinions and to try to convince me otherwise (but ultimately the final decision is with @trongnhanuit and @bqminh who might already disagree with me).

-major obstacle & open question:
There is no statistical difference between a fasta multiple sequence alignment and the maple format. They are two different representations of the same data, and you can losslessly convert between the two. So in your proposed procedure they would give exactly the same results. (I am ignoring the issue that when aligning many genomes we usually align them to a reference and discard insertions with respect to the reference; this affects both fasta and maple format alignments, so while it is an issue, it is not relevant here.)

Best wishes,
Nicola