validating identified sub-populations of cells in scRNA-seq

Question

In the analyses of single-cell RNA-seq data there are different unsupervised approaches to identify putative subpopulations (e.g. as available with Suerat or SCDE packages).

Is there a good way of computationally validating the cluster solutions? Different methods may results in slightly different clustering results. How to know which one is the best i.e. representative of biological sub-populations?

There's a bit of folly in trying to computationally validate computational methods without a known truth dataset... — Devon Ryan, Jun 02 '17 at 19:29
I changed scRNA to single-cell RNA (since scRNA also means small conditional RNA). — Kamil S Jaron, Jun 02 '17 at 19:39
@Kamil “scRNA-seq” is an established name for single-cell RNA-seq. — Konrad Rudolph, Jun 03 '17 at 12:33

score 6 · Accepted Answer · answered Jun 02 '17 at 19:50

A SC3, single-cell consensus clustering, approach could be helpful here. It aims at achieving "high accuracy and robustness by combining multiple clustering solutions through a consensus approach" https://www.nature.com/nmeth/journal/v14/n5/full/nmeth.4236.html

score 3 · Answer 2 · answered Jun 14 '17 at 01:38

While better methods of evaluating your clusters would be to use an external dataset or a dataset with known truth, there are a variety of internal validation metrics that can be used to compare clustering solutions without another dataset.

Here are a few metrics:

Davies-Bouldin Index
Calinski-Harabasz Index
Root-Mean-Square Standard Deviation

Many more can be found in this clustering review: http://stke.sciencemag.org/content/9/432/re6

These internal validation metrics grade your clustering solution based on three measures: compactness, connectedness, and separation. When using these metrics to compare clustering solutions, be sure to consider which metric is appropriate for your results as some algorithms work by optimizing certain measures.

validating identified sub-populations of cells in scRNA-seq

2 Answers2