Clarification on the use of delta variance

[juliaszusz]

Hi, thanks so much for putting this package together. It is such an important issue in single cell analysis that you have highlighted in your paper. I wanted to ask for clarification on your suggested application of the delta variance calculation.

Do you have a suggested method for determining what you consider to be a high value of delta variance? i.e. Do you suggest that above a certain numeric threshold value is considered high, or would it be more dataset dependent where you would suggest values above a certain percentile or distribution are considered high?

Also, I just wanted to clarify that the variance would be considered high just in terms of magnitude irrespective of the sign (so a large negative value is also considered high delta variance?)

I've applied this calculation to my dataset of ~27,000 genes from 19 individuals and have values ranging from -160180363 to 1259. The distribution is S shaped, with ~60% falling between -10 and 10, and approximately 35% of my genes have negative values below -10 including some extreme values, and around 5% of values are above 10.

Thanks so much again!

Julia

[jordansquair]

Hi Julia,

Sorry for the delay and thanks for your feedback.

I first just wanted to clarify that the use of delta variance is only relevant if you are using single-cell DE methods. If you are using pseudobulk methods then you don’t need to worry about this bias.

Like you said, the absolute values are going to vary depending on the depth of the dataset/normalization/method used. However, we also observe a similar sigmoidal relationship in most datasets. The best advice I can give is 1) to use pseudobulk methods out of the gate if it’s possible in your data (which it sounds like it is) to avoid this problem altogether, and 2) if you are hoping to validate any of the genes you identified in your DE analysis that fall in this bottom tail of the distribution I would really investigate their distributions manually to see if they look robust before proceeding with any wet-lab validation.

[Jorges1000]

Thanks for the great package and paper! I have trouble matching the output of calculate_delta_variance and run_de as the numbers of genes per cell type are different.

[jordansquair]

calculate_delta_variance returns a list, with a vector for each cell type. The genes should be the same as that from run_de. Can you maybe provide a sample of your data to my email (in my profile) as this isn't a bug (e.g., not a problem in the example dataset provided).