Congratttttttts!!!! So well deserved :)

If you want to mess around with some motifs, check out:
github.com/stephkoest/E...

It was a pleasure to work with you! 😊

Hey Felix, good question! Yeah it is different. In short: trimAl/ClipKit aim to remove uninformative sites. WitChi is a second step to remove misleading sites: those that can group unrelated taxa just because their sequence composition looks similar. Hope that helps!

Thanks, jolien! ;)

And of course the great work!

Thanks for that beautiful summary, kassi :)

It was a fun project :) Thanks for the support!

9. TL;DR + link dump
WitChi is:
✔ Fast
✔ Interpretable
✔ Tree- and model-free
✔ Benchmark-validated
Designed to fix compositional bias at phylogenomic scale.
With: @kassipan.bsky.social, @danieltamarit.bsky.social, @ettema.bsky.social
💻 github.com/stephkoest/w...
📄 www.biorxiv.org/content/10.1...

8.
GTDB r220 case study (led by @kassipan.bsky.social )
Applied WitChi to the archaeal GTDB r220 supermatrix:
• 5,869 taxa
• 55% of columns pruned
• Biased taxa: 95.1% → 2.3%
• Runtime: <2h on 4 cores
→ Known clades recovered — without using very complex C60 or CAT models

7.
Use witchi test to quantify bias per taxon:
• χ² scores
• Empirical p-values (via permutations)
• Z-scores to see how far taxa deviate from expectation
→ Great for screening MSAs or comparing compositional distortion across datasets.

6.
WitChi solves both problems:
🔹 Builds a null distribution using column permutations — no model, no tree
🔹 Recursively removes columns that distort the taxon-wise χ² profile
🎁 Bonus: 3 scoring strategies, including one capturing distribution-wide effects (Wasserstein)
⚡ Scales linearly with taxa

5.
Classical χ² pruning trims biased columns once — fast, but naive.
→ As alignment composition shifts, Δχ² must be updated — few tools do this.
BMGE’s stationary-based algorithm prunes iteratively and works well, but scales quadratically with taxa — not feasible for medium sized or large datasets.

4.
The problem:
χ² assumes taxa are independent and identically distributed samples.

In MSAs, they share history → correlated data.
So parametric χ² nulls are invalid.
Simulations help, but they need known models and trees — which bias distorts.
→ Slow, circular, rarely used.