I have two questions regarding the application of qiime2 to perform diversity analysis in soil samples. I am aware that “general solutions” are far from a realistic scenario but any comment coming from this great forum would be highly appreciated!
In general, typical studies of soil microbiomes include “noisy” situations like (among others): very high alpha diversity (thousands of ASVs in a single 0.25g sample), high natural betadiversity (considerable differences in microbial composition within samples of the same “treatments” or “biological replicates”), and of course uneven (> 10x) sample sizes due to complications with PCR amplifications in specific samples. Besides, often experimental designs are complex -not typical treatment vs. control situations- but might include several factors with interactions, and also random factors (block desingns). Taking all this into account, I sometimes hesitate if statistical methods that were primarily designed and tested for gut/human/clinical samples can be directly applied for environmental research. In this context:
Q1. About “normalization”: I just discovered the new q2-repeat-rarefy (GitHub - yxia0125/q2-repeat-rarefy). Thanks for the new tool! Repeating the rarefaction sounds conceptually a good idea with soil samples, considering that lots of low abundant and perhaps meaningful ASVs would disappear in a “single shot” rarefaction. Would you recommend this approach versus the “traditional” one (single rarefaction)? In that case, how can we use the output of q2-repeat-rarefy to feed the qiime diversity core-metrics plug in? (That I guess so far it performs single-rarefactions, right?).
Q2. About differential abundance analysis: I know this is a complicated topic with lots of discussions, but would you recommend (or directly discard) any of the available methods in the case of soil samples? If I understand correctly, both ANCOM and the most recent ANCOMBC (GitHub - FrederickHuangLin/ANCOMBC: Differential abundance (DA) analysis for microbial absolute abundance data) might be a good choose to look for “responsive” ASVs among groups of samples (say warm vs. control soil samples), but it is not possible to include interaction of factor terms (like temperature x humidity). Deseq2 allows those more complicated designs, but I read in this forum that it is not recommended for microbiome analysis anymore due to large numbers of false positive. I did not find any method that allows random effects.
I (personally) stay away from DeSeq2. Partially because I’m not an R user. Partially because I dont like the assumption they use to break compositionality for most studies (a normalized number of copies per enviroment.) You could look into a recent discussion around models with random effects and interaction terms:
It does require work outside of QIIME, but it might be an option. It can also be hard to interpret. (But so can ANCOM and any one of a number of other techniques)
Hello @jwdebelius ,
Thanks a lot for your ideas, really appreciated
I have check several soil-microbiome papers using diff abundance methods (deseq2, ANCOM, mixed lineal models and others are common). What is still shocking for me is I cant find many examples where authors applied at least two of these techniques to confirm their predictions. In reality, I guess this is not done because results differ among those techniques, especially in the case of complex samples/designs, and it complicates a lot the task of writing a paper. What makes me feel skeptical about the robustness of this part of the microbiome research.
In this recent paper (2020), authors recognize what I was speculating “Differences between these methods can seriously affect the biological interpretation of metabarcoding data, especially in ecosystems with high microbial diversity –soils-, as the methods are benchmarked based on low diversity datasets.” And moreover, based on simulated, mock, and real soil communities (16S and ITS) they do not recommend going beyond family level when applying diff abundance methods (!).
Perhaps we are simply pushing the techniques towards the limits. Bioinformatics allow us to decompose sequence data in fine-grained level (ASVs) but statistical testing is not enough developed yet for certain tests. But it is just my impression.
So we still need to learn a lot more from future research
I’ve applied multiple techniques to datasets. The ones that have high FDR occasionally detect taxa identified by another method. (A broken clock is right twice a day . But, I’m not sure “we know the method can be incorrect” is a good reason to justify that.) You also have the rapid changes in the field and work to adopt prevailing norms. Figuring out what the right method is can be really hard, since most developers want you to use their method. (Morton et al did compare their method in a moderate complexity enviroment and argue its better)
When I read the paper you linked, though, it looks like they haven’t fully exorcised other differential abundance methods, but rather explored the whole pipeline. (I saw a negative binomial with a GLM, but skimming through their method and their code, it doesnt look like they compared a lot of differential abundance methods and that they mostly focused on their OTU/ASV comparison.
The last big comparison paper I saw across multiple differential abundance methods was based on R, and missed a lot of the newer ALR/IRL methods, maybe because it took two years to publish. (And I can’t find it right now! ).
I think we’re also asking a lot of a few field with very complex data and complex data constraints. Many of the differential abundance techniques do struggle with complex experimental designs. Others can handle complex models but are harder to interpret, which makes people shy away from them. (IRL and reference frames are great for complex models but get harder and harder to interpret.)
The “best way” to handle differential abundance has also changed approximately every two years throughout my career (kruskal wallis → DeSeq2 (2014) → ANCOM (2015/2016) → ILR (2016/2017) → Reference frames/ALR (2019) → Bayesian magic (predicted 2021)) and there are constantly new differential abundance methods coming out that promise to be better. There are also different methods implemented different places. Which major library a package interfaces with definitely influences who uses it and how.
Plus, people are still constantly going back to older methods (i.e. LefSe, DeSeq2) because they’re well cited, the user is familiar with them, or because they dont necessarily make the best assumptions about the data, they tend to give more differentially abundant results. (My cynical view is that people often feel like they haven’t had the full microbiome experience if they don’t get differentially abundant taxa to talk about.)
So, overall, I think you’re right that we have current statistical limitations. It sometimes feels like we’re trying to do rocket surgery with knives we chiseled from obsidian and the light of a fire flickering in the wind.