Seurat integration can improve and deteriorate clustering

[alexvpickering]

Here is the original labels for the healthy and diseased SJIA lung samples but with TSNE coordinates after integration:

And here are the clusters I labelled after integration:

For the most part, the Seurat integrations seems to improve the clustering of the samples. For example, some diseased cells previously labelled as B-cells seem better labelled as RBCs (note that expression values are from pre-integration):

In contrast, I believe that the distinct Alveolar Epithelium clusters is an artefact of there being no Smooth Muscle cells in the healthy sample and the integration algorithm matching the healthy Alveolar Epithelium cluster with the diseased Smooth Muscle cluster. There are some top marker genes that share expression between healthy and control cells within this cluster:

There are other top marker genes that are very distinct between healthy and diseased cells in this Alveolar Epithelium cluster:

Additionally, if the Smooth Muscle cluster is excluded prior to integration, the Alveolar Epithlium cells do not form two distinct clusters after integration. Here is the full PDF for the combined reports:

lung_combined_markers.pdf

And for documentation, commits related to generating these reports include:

• 4257919dcfbc7a76027483bc032d7e9567f46d17
• 2b103c1e99af3f68ecee2124322e2de64fc7c00b
• 7a5e29b06d9db314f52101995f8960d221482c6b
• 902e8d4d7538dd0a9db303d00022d367c228d71f