Choice of reference and query data sets

[hl-xue]

Hello,

Thanks for the nice software.

I have 2 data sets A and B annotated separately, and I am using scArches for a label transfer analysis from one data set to the other to identify their potential similarity and difference.
I am following the tutorial here except for skipping the step 4a, in order to compare predicted cell types from the other data set with the actual annotations.

However, I found that different choices of reference/query data sets make remarkable difference.
For example:

• when I use data set A as reference and data set B for query, I got a normalised MI score as 0.71;
• however, when I reverse the order using data set B as reference and data set A for query, the normalised MI score turns to be 0.79.
• The confusions of cell types between two versions also look quite different.

In this situation, I would like verify with you about two questions:

1. Is this difference in label transfer results caused by reversing order of reference and query data sets is expected? If so, why could this happen?
2. If the difference is expected, which result should I trust?

Thanks in advance!

[hl-xue]

Hello,

I have a further question please: when I try to label transfer with the same workflow from one data set to itself (use data set A as ref and the same data set A as target to predict). I got non-perfect matching (with normalised MI score as 0.89). Is this also expected?

Thanks again!