conclusion.Rmd

# Conclusion

It is commonly assumed that differential abundance (DA) analysis is robust to taxonomic bias, so long as all samples have been subjected to the same MGS workflow.
In contrast, our results show that consistent taxonomic bias can distort the results of DA methods that are based on species proportions.
This distortion occurs because the fold error in a species' proportion depends on the sample mean efficiency, which can vary across samples and conditions in accordance with the overall community composition.
This problem continues to apply for analyses of absolute abundance that are based on multiplying the proportions from MGS by the total abundance of the community.
In contrast, DA methods based on species ratios are invariant to consistent bias, since the error in the ratio between species is independent of community composition.
In Section \@ref(solutions), we describe how ratio-based methods can thus provide a more robust approach to analyzing relative and absolute abundances;
in addition, calibration with community controls, bias-sensitivity analysis, and bias-aware meta-analysis provide additional ways to correct or mitigate the effect of bias.

Proportion-based DA methods encompass many of the most popular methods for analyzing relative and absolute abundances.
Importantly, however, if the mean efficiency is approximately constant across samples, then bias has a negligible effect on multiplicative and rank-based DA results.
If the mean efficiency varies but is unassociated with the condition of interest, then bias merely serves to increase noise and does not create systematic errors in DA results.
It is only when the mean efficiency is associated with the condition being analyzed that large systematic errors can occur.
Our case studies suggest that this problematic scenario does occur, but it may be the exception rather than the rule.
Systematic investigation of how the mean efficiency affects DA results across a wide range of studies may increase our confidence in previous DA results and/or alert us to the conditions in which they are most suspect.

Important open questions include 1) determining the extent to which bias is consistent across samples for different taxonomic levels, MGS methods, and sample types; 2) assessing the validity of post-extraction measurements as measures of pre-extraction abundance; 3) understanding how the multiplicative error from bias interacts with the non-multiplicative error from contamination and taxonomic misassignment; and 4) understanding how different underlying community dynamics, and in particular the source of zero counts in MGS measurements, affect bias sensitivity.
In addition, while we have showed some simple methods for incorporating bias and control measurements into DA analyses, more sophisticated statistical tools are needed to properly account for both taxonomic bias and random variation in the underlying MGS and supplemental (e.g. qPCR) measurements.
Finally, more concrete experimental recommendations and user-friendly statistical workflows are needed for implementing the solutions we propose in various experimental contexts.

Our theoretical framework and example analyses provide a foundation for addressing these open questions, as well as developing experimental protocols and statistical tools that implement our proposed solutions.
We look to a future in which microbiome researchers regularly choose an appropriate combination of experimental and data-analytic methods that are capable of answering their fundamental question while also accounting for taxonomic bias and other limitations inherent to MGS measurement.
In doing so, we will collectively gain the confidence needed to codify the findings from MGS-based microbiome studies into true scientific knowledge.