Use of FlashWeave or FlashWeaveHE

[YibiChen]

Hello, thanks for developing this useful tool. I have a question regarding the choice of FlashWeave versus FlashWeaveHE.

I have around 500 samples and 5000 OTUs from three habitats. I am pretty sure many of these OTUs are habitats specific. I am wondering which mode should I use. From what I understand, FlashWeaveHE is to acount for structural zeros, which lead to false-positive OTU-OTU associations. However, if I put habitat in meta-variables, wouldn't these false-positive associations be identified as indirect associations?

If this is the case, perhaps I should disable the heterogeneous mode? I imagine heterogeneous mode would require more samples. Also, any recommendations for the smallest sample size?

[jtackm]

Hi Yibi,

Sorry for the late reply. Your assertions are correct: FlashWeaveHE more thoroughly removes the effect of structural zeros in large data sets, but regular FlashWeave (in particular when provided with meta variables) can also reduces these effects in small to medium sized data. Sample limits depend on the dataset, but 500 samples is on the low end for FlashWeaveHE. I'd consider running regular FlashWeave (the default option, heterogeneous=false in learn_network()) in your case, and when in doubt comparing it to FlashWeaveHE later - knowing that missing links may be either due to structural zeros or to power issues.