Hello,
I heard that there is a new dimension reduction technique that for visualization, and it better than UMAP and TSNE. However, anyone heard about this new technique? And how is it performance?
Thanks
Andy
Many people can (and do) say that their dimensionality reduction technique is better than all others, and come up with data and benchmarks to support that claim. Up until recently, it was a popular opinion that UMAP is much better than t-SNE for analyzing scRNA-seq data. Then this paper came out showing that this supposedly superior behavior of UMAP is mostly eliminated by applying similar initialization procedures with t-SNE. I say mostly because UMAP is still faster and often creates tighter groupings of data points than t-SNE, which may be desirable to many people. Yet I don't think UMAP is in any fundamental way a better choice than t-SNE, even though by now that may be a prevailing opinion.
This is a long-winded way of saying that you may want to consider not trusting all the way what someone advertised about better dimensionality reduction methods.
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
This is an active area of research. One method that comes to mind is PHATE (paper 1, paper 2), but it's not the only one.
Thank you for your reply. I found another one as well, which is called "deep-tda." Unfortunately, this one does not have open resources.
The one you shared seems worth trying. I will give it a try. Thanks for your help.
There are many new dimension reduction techniques being published, e.g. PHATE https://www.nature.com/articles/s41587-019-0336-3
Whether they are better than UMAP and t-SNE: Just read the papers with their benchmarks.
I'd wager that they probably aren't much better -- https://www.biorxiv.org/content/10.1101/2021.08.25.457696v4.full -- these techniques inevitably "butcher" your data and analysis is better conducted in the ambient space. UMAP/t-SNE-type stuff is mostly useful if you want to see points colored by cluster on a 2D plot, keeping in mind that the distances between points shouldn't really be interpreted.
Thank you for sharing the link to the PHATE dimension reduction technique article.
FYI, the package I mentioned is "deep-tda" as a package which I heard about from my college. It's unfortunate that the package is not open for use.
I am not the creator of deep-tda. I simply heard about it from someone in a neighboring lab. However, the person I spoke to didn't provide clear information, which is why I asked the question in the first place. Just a few minutes ago, I had the opportunity to meet with him again and asked for the name once more.
Yes, I read about that on twitter, and the responses were critical of its lack of open benchmarking, lack of open source code, lack of explanation of how it works, among many things. That raises red flags. Would not use it (again, I think try to avoid these arbitrary, distorting, dimensionality reduction techniques in general if possible -- just look at your Leiden clusters!).