martinjzhang
commented
2 years ago Hi,
Thank you for your interest in our work! This is indeed an expected behavior.
Specifically, the scDRS results of a given cell depend on other cells in the data set through both the estimation of technical noise levels and the selection of matched control genes (expression mean and variance are evaluated across all cells in the data). In a way, scDRS assesses if a given cell has excess disease enrichment compared to an "average cell", and this "average cell" is defined by other cells in the data set. As a result, including only causal cells will likely yield non-significant results.
To see this, consider the case where MSN is the causal cell type and all putative disease genes are highly expressed in MSNs (and lowly expressed in other cells). If we include only MSNs, the mean expression of the disease genes will be very high. As a result, the matched control genes will also have high expression, producing high control scores and making the disease scores non-significant. However, such a situation would not appear if we include all cells in the TMS data, where the mean expression of the putative disease genes, averaged across both the causal MSNs and other non-causal cells, will not be very high.
Please see also the discussion points 6 and 7 in the manuscript (lines 472-481; https://www.biorxiv.org/content/10.1101/2021.09.24.461597v1.full.pdf).
We recommend including all cells in your scRNA-seq data set. In general, to run scDRS for a given disease gene set, the scRNA-seq data should ideally contain a good proportion of non-causal cells to provide enough contrast to the causal cell population.
Let me know if you have more questions.
Best, Martin
NoahMottelson
Hey,
Playing around with reproducing your results, I noticed that results for a given set of cells depend heavily on what other cells are included in the analysis. For example in the TMS dataset, the medium spiny neurons came out highly significant for several traits when analysed in the context of the entire dataset (as you show), but insignificant when analysed on their own. I was surprised by this because, to my understanding, the background cells are only used to normalize the score via their control scores,which don't have any biological interpretation.
Is this the expected behavior? If so, do you have any recommendations regarding the number and diversity of cells to include in an analysis, above and beyond the cell population of interest?