Tue Jun 23, 2015 6:27 am by Evan
Hi Matthew,
Thanks for your comment. For both of these papers, our goal was to better understand the composition of the cell populations. I agree that tSNE does a very good job at delineating sub-populations in a visually verifiable way, even when some are quite rare. Nonetheless, some populations of interest such as Th2 or other rare helper subsets do not always nicely segregate. This is probably because they are too rare and/or because there aren’t enough other defining features (e.g., consistent expression or other markers being used – or they are also heterogeneous, etc..). Furthermore, as you point out, there are tradeoffs in terms of optimizing signal for all the markers we are using. So, for some markers, we still have some room for improvement. We included examples of each of the stains within this dataset in the supplementary material to illustrate these limitations. Although better staining may help to improve the tSNE plots and the resolution of these populations, we are satisfied with our ability to set objective and mostly accurate thresholds for defining cells that are positive (and high/low in some cases) vs. negative for each of the markers based on these plots and on quite clean plots of unstimulated cells we ran using the same panel.
In terms of clustering algorithms, we agree with your comment and used densVM because it performed better (visually) than other methods we compared. However, I think that there is room for improvement here as well in that some small but visually discernible clusters are not always identified and other large amorphous clusters seem to be arbitrarily segregated. I think that continued evaluation and improvement of clustering algorithms would be useful in this respect. I also think that manually defined cluster boundaries (taking advantage of the visual nature of dimensionality reduction methods!) are also useful and usually more accurate at this stage. That is, if using subjective gates can be accommodated in the analysis, I think it makes good sense to just gate on them by hand. I’d argue that clustering algorithms and choices of parameters for these algorithms are always going to be subjective anyway - but of course it depends on the purpose.
Cheers!
Evan and Mike W.