Cleaning up cells in clusters that clearly do not group well.

Question

[kenneditodd]

What analysis can I run that will help remove abberant cells? or cells that deviate too far from the heart of the cluster? For example, there are 4 cells from cluster 1 that are all over the place and far from most cluster 1 cells. I want to 'clean' this up but obviously in an unbiased way to get rid of outliers. What analysis do you recommend running for removal?

[LauferVA]

please state your intended work environment and software choice(s)

Answer 1

[Mensur Dlakic]

If you do density-based clustering, for example using the DBSCAN algorithm, the outliers will be automatically removed from well-defined clusters. Below are links with an explanation and with two different python implementations.

• https://towardsdatascience.com/dbscan-clustering-explained-97556a2ad556
• https://hdbscan.readthedocs.io/en/latest/index.html
• https://scikit-learn.org/stable/modules/generated/sklearn.cluster.DBSCAN.html