Differential Abundance

Proportionality of scRNAseq data

These points are taken from the tutorial for scCODA, a scRNAseq-dedicated differential abundance (DA) tool.

• 10x single-cell experiments are limited by their throughout: a maximum of 10k cells can be assayed.
• This means all abundances are proportional: an increasing of cell A may instead be due to a decrease of cell B, etc
• This is ultimately unavoidable, and while scCODA boasts a Bayesian statistical framework to counteract the impact of inherently proportional data, the best solution may simply be a parallel flow cytometry experiment (as was performed here).

Cell-free DA: milo

• milo seeks to identify groups of DA cells in a cluster-free manner (or, within clusters).
• It does this by performing DA analysis directly on the neighbourhood graph, identifying DA neighbourhoods (each composed of roughly 50-100 cells). This can be thought of as performing a highly granular over-clustering, then performing normal DA analysis.
• milo also allows probing DA neighbourhoods expression trends.
• milo is implemented in R and Python.

Importantly, milo works directly on the neighbourhood graph. As this can be calculated using a batch-corrected/integrated embeddings milo represents the best way to analyse granular differences in an integrated dataset (where DE is impossible).