Best practice for Milo with mixed-effects models and multiple comparisons

[nmdrf]

Hi,

I am using Milo with mixed-effects models and have a question about best practices when performing multiple pairwise comparisons.

When testing several contrasts between experimental conditions, there seem to be two possible approaches:

1. Build the kNN graph on the full dataset
   • Construct the graph and neighborhoods using all conditions.
   • Run differential abundance testing for each contrast separately.

2. Build the graph separately for each comparison
   • Subset the dataset to only the conditions involved in a given contrast.
   • Construct the graph and neighborhoods on that subset before running the test.

In my analysis, I observed that these two approaches produce noticeably different results. When building the graph on the full dataset, the results for specific contrasts become less interpretable, whereas building the graph separately for each comparison produces results that appear more consistent with expectations.

My questions are:
• Is there a recommended best practice for this scenario?
• Should the neighborhood graph ideally be built globally across all conditions, or separately for each comparison?
• How is this situation conceptually handled within the Milo framework?

Any guidance or experience with this type of analysis would be greatly appreciated.

Thanks!

[MikeDMorgan]

At the end of the day, if your cell types are very well separated, and biologically there is no expected lineage relationships between them, then splitting them out will give much higher resolution of cell states than trying to create one global analysis.

For different conditions, this is entirely up to you and how you wish to report any comparisons. You should note that because the NB-GLM used edgeR, it benefits from more samples to estimate the biological coefficient of variation, and hence reducing the estimation error.