# CTF with interventions

**URL:** https://forum.qiime2.org/t/ctf-with-interventions/20775
**Category:** Library Support
**Tags:** longitudinal, intervention-study, gemelli
**Created:** [September 16, 2021, 1:13pm UTC](https://forum.qiime2.org/t/ctf-with-interventions/20775 "2021-09-16T13:13:24Z")
**Posts on this page:** 2
**Page:** 1

<div class="post-metadata">

### Author: ![jwdebelius](https://forum.qiime2.org/user_avatar/forum.qiime2.org/jwdebelius/32/9655_2.png) [@jwdebelius](https://forum.qiime2.org/u/jwdebelius)
#### Post date: [September 16, 2021, 1:13pm UTC](https://forum.qiime2.org/t/ctf-with-interventions/20775/1 "2021-09-16T13:13:24Z")

</div>

Hi (probably) @cmartino or other appropriate person,

I know at one point you mentioned that it's gradient-aware but not longitudinal. In the examples from both the [original paper](https://www.nature.com/articles/s41587-020-0660-7), [a study of the built enviroment of SARS-CoV2](https://www.ncbi.nlm.nih.gov/labs/pmc/articles/PMC7685343/) by Rob Knight's group and [an infant transplantation study](https://www.sciencedirect.com/science/article/pii/S2666634021002038), it looks like Gemelli was applied to look at dyanmics _after_ the intervention/grouping. (i.e. comparing comparing vaginal/cesarean born-infants; IBD patients vs healthy controls).

Is this the recommend use case, rather than say, working with something where there's a baseline sample followed by an intervention?

Best,  
Justine

---

<div class="post-metadata">

### Author: ![cmartino](https://forum.qiime2.org/user_avatar/forum.qiime2.org/cmartino/32/3700_2.png) [@cmartino](https://forum.qiime2.org/u/cmartino)
#### Post date: [September 16, 2021, 3:54pm UTC](https://forum.qiime2.org/t/ctf-with-interventions/20775/2 "2021-09-16T15:54:46Z")

</div>

Hi @jwdebelius,

Good question, thanks for asking! I think the best way to describe [gemelli](https://github.com/biocore/gemelli) / [CTF](https://www.nature.com/articles/s41587-020-0660-7) is that allows for context-dependent or repeated-measure data dimensionality reduction. It is not restricted to longitudinal repeated-measure data because the order does not matter but that also means it is not explicitly modeling time. In that way, it is truly an unsupervised dimensionality reduction (like PCoA on beta-diversity distances). I will note that it is not unbreakable, so filter or structure the data accordingly. For example making sure the data is well balanced, remove a time-point where only one sample was collected, or remove a subject with only one time-point.

Does that help bring some clarity to the method?

Thanks,

Cameron