Ohh curious to see your work too then :)

MapQC is a pip-installable python package, and runs in less than 2 mins on a query dataset of 30k and a ref. of 0.5M cells. For more info, see our GitHub repo github.com/theislab/mapqc. It also includes tutorials. Try it out, and let me know what you think!

Re-mapping with different parameters drastically improves the mapping, and enables identifying disease-specific cell states, such as altered smooth muscle cells in lungs of patients with IPF (reproducible across studies). MapQC thus helps you make the best use of your data!

Here’s an example of a mapping that failed. For some cell types (e.g. circled ct), the UMAP suggests the query mixes well with the reference. However, mapQC scores show this is not the case, and downstream-analysis indeed results in batch-effect driven conclusions.

This results in cell-level mapQC scores, with a score >2 indicating large distance to the reference. We expect controls in the query to be the same as in the ref., i.e. to show scores <2. In contrast, disease samples should show some high distance to the ref. (local scores >2).

How does it work? We use the control samples in the large-scale reference to obtain prior knowledge of normal inter-sample variation. We do this locally, such that we learn cell-state-specific inter-sample distances. We compare those to query sample distances to the reference.

We therefore developed mapQC, a method that takes as its input any query mapping to a large-scale reference, and outputs a cell-level mapQC score. The score will tell you if, and where, the mapped query contains batch effects, or if e.g. disease-specific variation was removed.

One commonly used metric, LISI, is highly sensitive to cell numbers, which in fact are independent of integration/mapping quality and should not affect metric outcome. Finally, all of the existing metrics lack a clear rationale for a cutoff between good and bad mappings.

Moreover, standard integration and mapping metrics fail to pinpoint these failures: they quantify the wrong things. Here’s an example, with one very poor and one good-quality mapping resulting in the same scores for several metrics, but not mapQC.

With the surge in large-scale single-cell atlases, many people have started using atlases to analyze their new data. However, query-to-reference mappings, used to combine a reference with new data, often do not produce a good embedding. This leads to data misinterpretation.

7/7 We hope that our guide will help you to construct high-quality atlases and efficiently explore their contents. We would love to hear your thoughts on atlasing and your comments on our guide!

6/7 When building atlases, the downstream use-cases should always be the primary focus. Thus, we extensively discuss how atlases can be used in single-cell and broader biological research.

5/7 Atlas building involves many steps: defining the atlas’ focus, data preprocessing, integration, annotation and evaluation. Afterwards the atlas must be shared with the community and eventually updated and extended. We present diverse considerations associated with each step.

4/7 However, constructing an atlas is not straightforward and the methods in the field are still evolving. With this in mind, we discuss different approaches for atlas building, along with their pros and cons. This will help you make better informed decisions in upcoming atlasing projects.

3/7 Atlases have value beyond individual datasets. Their diversity in samples, donors and conditions paint a more holistic picture of biology. Moreover, they are invaluable as a reference for analyzing new data, easing preprocessing and guiding interpretation.

2/7 The surge in single-cell datasets improved our understanding of biology, and integrating these datasets into unified “atlases” can teach us even more: we can create consensus cell type naming, increase power for learning disease-related patterns, and compare across multiple diseases.