MVP demonstration on lung data

[emdann]

Lung dataset location: /home/jupyter/mount/gdrive/query2reference/processed_data/lung.h5ad

With celltypist/celltypes

• Simulate cond-ctrl-ref design with simulations (see ctrlSimulations)
• Train celltypist model on reference
• Predict labels on control and query
• Differential uncertainty test (_on majorityvoting or not?)
• Does it detect true query-specific population? Using intersection/NMI between predicted clusters and true clusters (is the clustering choice problematic here? when using majority voting)

With scArches/nhoods

• Simulate cond-ctrl-ref design with simulations (see ctrlSimulations)
• Train scvi model on reference
• Predict labels on control and query
• Differential uncertainty test on nhoods (milopy defined)
• Does it detect true query-specific population? Same metrics used in milo benchmarking

[emdann]

Here we don't have enough replicates to do the differential analysis and use the rest as reference. Use the integrated PBMC datasets instead.

[emdann]

Or use full lung cell atlas https://cellxgene.cziscience.com/collections/6f6d381a-7701-4781-935c-db10d30de293

[emdann]

• Does the test on uncertainty detect the query-specific population?
• Does it detect it more sensitively than differential abundance or the single metric?

[emdann]

Answer: no