Thread by @flo_compbio

If you’re interested in clustering single-cell RNA-Seq data, please check out my new paper “Straightforward clustering of single-cell RNA-Seq data with t-SNE and DBSCAN”! biorxiv.org/content/10.110…

Working PBMC example: github.com/flo-compbio/ga…

Quick Twitter summary below: 1/

I propose “Galapagos”, a clustering workflow designed to reliably separate the main cell types in scRNA-Seq data in a straightforward and transparent manner. 2/

The workflow consists of a simple series of steps and does not require gene selection. Essentially, clustering is performed by applying DBSCAN directly to t-SNE results. 3/

As there is some concern as to whether t-SNE is sufficiently robust for clustering purposes, I show that when t-SNE is applied to variance-stabilized data (as described in the paper), the results are indeed highly robust to parameter settings and initialization points. 4/

I then adopt a simulation approach to test whether Galapagos is able to overcome the high levels of technical noise in scRNA-Seq data and generate accurate clustering results. 5/

The results on the simulated data closely match the true cell type labels obtained from the “ground truth”, suggesting that Galapagos results are indeed accurate. 6/

To quantify clustering accuracy based on an experimental ground truth, I use CITE-Seq data to compare Galapagos results to cell type identities established based on protein expression markers. 7/

The results show that the accuracy (precision & recall) is >90% for the three cell types examined.
8/

Finally, I again use the CITE-Seq data to experimentally define different subsets of T cells. A comparison with Galapagos results suggest that Galapagos is able to distinguish between naive T cells, CD4+ memory T cells, and CD8+ memory T cells. 9/

However, distinguishing between CD4+ and CD8+ naive T cells is nearly impossible, as t-SNE fails to clearly separate these populations. 10/

In conclusion, Galapagos represents a straightforward approach for clustering single-cell RNA-Seq data that can be implemented in most programming languages with only a few lines of code. 11/

Since clustering is performed directly on the t-SNE results, the fine-tuning of the two DBSCAN parameters is very intuitive (“Do the clusters match my interpretation of the t-SNE plot?”) 12/

Clustering on the t-SNE results also makes the clustering results very easy to communicate. (“What you see is what you get.”) 13/

In the Discussion, I argue that the problems with t-SNE have more to do with the fact that there are so many different ways that the data are filtered and transformed *prior* to the application of t-SNE. In Galapagos, I therefore tried to simplify and standardize these steps. 14/

The main limitation of the method as I see it is the difficulty to distinguish between closely related cell types in cases where t-SNE fails to clearly separate the corresponding cell populations. 15/

In this case, my suggestion would be to apply denoising, and apply clustering to heatmaps of highly variable genes. (For more on this point, see our ENHANCE paper: biorxiv.org/content/10.110…) 16/

Please let me know if you have questions or suggestions for improvements. If you are potentially interested to serve as a reviewer for this paper for an open-access journal (e.g., F1000Research), please let me know as well. Thank you! (end of thread)