I mean if I was the other guy, I'd have articulated a criticism something like:

> Using parsimony to infer lineage seems a bit arbitrary since the constraints phylip pars uses in its clustering algorithm are intended for dna/rna/assays from organisms that have undergone evolution. And the over-represented words that rise to the top in a model's output aren't present/absent because of these same evolutionary dynamics. Also a model can have multiple "parents" whose outputs it was trained on, which would need a more complex representation of lineage than a dendrogram or phylo tree can show.

To which I'd reply something like:

The usage of the parsimony algorithm to infer the tree is defensible *if* there is signal indicating lineage in the raw data that isn't otherwise extracted by normal hierarchical clustering. For instance, phylip pars weights rare shared features more highly. If our data encodes signal of lineage in ways that somewhat align with the biological assumptions the parsimony algo is based on, it can get us somewhere closer to the true lineage, compared to hierarchical clustering. On the other hand, it might get us *further* from the true lineage if the parsimony constraints fixate on spurious signal, given that we're feeding it cross domain data.

The upshot of being wrong about this hunch that there might be signal that parsimony can pull out about lineage is simply that it behaves more like a naive clustering algo, perhaps producing slightly different trees. In practice, the trees generated with either method are very similar, though with a few interesting differences!

Since there's no way for us to validate whether one clustering method produces a tree closer to ground truth, other than the sniff test, I simply make no claims about *lineage* and present the charts as indicative of *similarity of slop profiles*. The strongest thing I will say as an interpretation is to speculate that their relatedness on the dendrogram may be indicative of which lab made the model or which models seeded its training data. Which I think is defensible regardless of which clustering algorithm is chosen, as long as I've been clear that interpretations like this are speculative.

One clear downside to my approach is that we lose a representation of similarity/distance which is normally shown via branch length when doing hierarchical clustering on similarity. I'm looking into fixing that.

The other clear limitation of this representation is that models can have multiple direct ancestors contributing to its training data, and our dendrograms collapse it to just one. But this critique applies to any clustering method that produces trees like this. To do it properly we could use network clustering or somesuch, though this is much less readable/interpretable.

So that's my hypothetical rebuttal to myself. Just to show that some thought actually goes into the methodological choices.

(I'm responding to you because I think the other person was just complaining to complain)