Why Does the `integration_consistency` require raw counts?

[ilan-gold]

See https://github.com/gao-lab/GLUE/blob/e20518b0ae7b39d47244087b825511e5c84f9b7e/scglue/data.py#L609 for the line from which integration_consistency makes this requirement. I am wondering why this is the case? I have floating point data I am trying to integrate.

[Jeff1995]

Well this situation a bit awkward. The integration_consistency involves computing feature-feature correlation. For the most common RNA & ATAC data, we need to preprocess the count data in a particular way (total count normalization + log transformation) to ensure that the correlations make sense.

For other data types it is not as clear what preprocessing is necessary. I guess the preprocessing part should ideally be modularized so users can also insert their own preprocessing function. I'll try to add that in the next release!

[Jeff1995]

Sorry for the late response!

Starting from v0.3.0, scglue.models.integration_consistency now works with non raw-count data. Nevertheless, you may need to specify appropriate "metacell" aggregation and preprocessing functions via arguments "agg_fns" and "prep_fns" (passed along to scglue.data.metacell_corr).

By default, the function retains the previous behavior of raw-count normalization if the datasets were configured with "NB" or "ZINB" models.