NAT Vs T samples DA analysis

Question

[San]

Hello, I am using a study's data which has NAT and T samples. My metadata is mixed that is I have both paired patient( patint w both NAT AND T samples) and unpaired i.e either NAT or T samples from a patient in my clinical metadata.

If i want to use these 4 DA tools (Deseq2, Aldex2, Ancom and Maaslin2) and make the analysis comparable with sig being padj<0.05, how can i make the deseign since I can only add Patient ID as random effect for Maaslin2 and Ancom, but idk how to model it for the rest of the two tools

I tried subsuetting and using the 4 tools and only deseq gave me sigf outputs and other tools had no <0.05 outputs. But maybe i am doing sth wrong.

My question is if i proceed using the 4 tools and not adding it as effect for some tools which do not accept it, I may not have a comparable output since different tools have different designs.

This is my first time working with microbiome data. so idk how to handle a clincial metadata with paired and unpaired patients.

Thank you

[Kevin Blighe]

What's your N for paired patients (those with both NAT & T samples)? It's crucial for tweaking the models:

• If low (<10-15), power's limited; ANCOM/Maaslin2 might stay conservative (few sigs at padj<0.05), while DESeq2 (using fixed effects) could overstate hits. Consider boosting with covariates or sensitivity runs.
• If higher (>30), random effects across all tools will balance things nicely—e.g., (1|patient_id) in ANCOM/Maaslin2, fixed ~patient_id + condition in DESeq2.

With N=20ish, you'd expect 5-20% sig taxa if effects aren't tiny. Share yours (and maybe total samples/depth)? I can paste-ready code snippets. First-time microbiome?

[San]

Thank you for your reply, Kevin!

Total samples: 506

Unique patients: 351

Paired patients: 154 , these have both Tumor and NAT (Normal Adjacent Tumor) samples

Unpaired patients: 197, these have only one group (either Tumor or NAT)

For differential abundance (species-level, after 10% prevalence filtering), I’m using four tools with the following designs:

MaAsLin2: abundance ~ Group + Center + (1 | Patient_ID)

ANCOM-BC2: ~ Group + Center + (1 | Patient_ID)

DESeq2: ~ Group + Center

ALDEx2: ~ Group + Center

Across 81 taxa, I’m getting fewer than 10 significant taxa (padj < 0.05) from most tools, and sometimes only 1–2. I’m wondering if the design might be too conservative or not appropriate for this type of mixed metadata.

I initially thought about using a fixed-effect design in DESeq2, like ~ Patient_ID + Group, but since over half my patients are unpaired, that leads to a non–full-rank model.

Would it make more sense to:

Restrict DA analysis to paired patients only (so I can use ~ Patient_ID + Group consistently across tools that allow pairing), OR treat all sample as independed and ignore the pairing? or

Keep the full mixed dataset and accept that only some tools (MaAsLin2 / ANCOM-BC2) can model Patient_ID as a random effect while others can’t?

I’d really appreciate your thoughts on what’s generally considered best practice for differential abundance analysis in mixed datasets like this (partly paired, partly unpaired).