Unexpected Projected Prediction and Log Ratio Plots from Gneiss

[Zachary_Bendiks]

Hello QIIME2 community,

I have microbiome data from a diet intervention study using a 3x3-factorial design - fat, grain, and starch. I'm using gneiss to associate groups of ASVs with host metadata using qiime 2018.4. I used correlation clustering to group ASVs and performed ols-regression with the formula "Cecal_pH+Abdominal_fat+Acetate+Butyrate+Propionate". I have some dataset-specific questions I'd like feedback on before I explore these results in-depth.

1. My understanding is that gneiss performs its own under-the-hood normalization of the feature counts. I therefore added pseudocounts and performed correlation clustering with the unrarefied ASV table as done here: QIIME2 Workflow. Is this thinking correct, or should I repeat the analysis with rarefied data?

2. I expect the measurements in my model to show at least some collinearity (ex: a drop in Cecal_pH is probably driven by increased acetate, propionate, and/or butyrate in this dataset). Will this affect how I build my regression formula?

3. In the attached regression summary, raw values in the "projected prediction" plot form three distinct clusters which most of the predicted values fall within. However, a minority of them fall outside the bottom-right cluster. The distribution of raw values looks strange compared to what I've seen elsewhere, and the presence of predicted values outside the "range" of raw values is troubling me a bit. Does this mean that the model has poor predictive power? Can I 'trust' any results I get?
   regression_summary.qzv (7.4 MB)

4. I explored the relationship between balance 'y4' and butyrate. In the log-ratio:butyrate plot produced by balance-taxonomy, I expected the relationship to be linear but it actually looks logarithmic. Is this a 'reasonable' result, or does it indicate some underlying issues with the model?
   y4_taxa_summary_butyrate_L4.qzv (126.4 KB)

Thanks for any help you can offer! If anything needs clarification please let me know!

[mortonjt (Jamie Morton)]

Hi @Zachary_Bendiks note that we are currently deprecating these functionality in gneiss and recommend to instead look at songbird and aldex2

There have been a few other related posts on this, so feel free to read those

Interpretation of gneiss lme-regression output

Hi @cdl, note that we are in the middle deprecating LME regression in favor of other differential abundances methods such as songbird and aldex2. See this post as well: Linear Regression Summary in Gneiss Concerning your other questions, y1 is the name of the log ratio. See the original gneiss tutorial on this https://docs.qiime2.org/2019.10/tutorials/gneiss/ The coefficients are slopes (i.e. differences across groups) in units of log-fold change (also covered in the tutorial). It’s not clea…

Some clarifications on gneiss correlation-clustering

Hi again @Sparkle, The visualizers you are using in gneiss are actually deprecated… when running them or reading the help docs you should see this note: This command is deprecated and will be removed in a future version of this plugin. So just beware that you are using a visualizer that the developers plan to remove in the next release of gneiss. The developer of gneiss, @mortonjt, now favors his newer method, songbird, so you may want to try that out instead: Because that ASV is only…

Regarding your questions

1. Yes, log ratios largely negate the need for rarefaction - but the issue with zeros remains. This is better handled in songbird and aldex2 due to their Multinomial model.
2. Definitely, collinearity is still a problem in differential abundance. The standard statistical recommendations still apply here - I'd start with fitting multiple models and comparing their differences.
3. We're deprecating this functionality and recommend songbird/aldex2 instead. So no, please don't trust those results.
4. Nature is often non-linear, so maybe this isn't completely unexpected. Linear methods can help with insights on this, but certainly won't be the best model for this.