chuckherring.bsky.social
Former US Marine, General Contractor, graduate student in Lau Lab, and current postdoc in the Lister Lab. Single cell genomics and epigenetics. Bioinformatics, and Neuronal Development 💻🧬🇦🇺
23 posts
77 followers
205 following
Regular Contributor
Conversation Starter
comment in response to
post
Transported for an instance to what it was like to read Carl Sagan many years ago. Thanks for sharing.
comment in response to
post
I started a construction business in 2007-08 (perfect timing 🫠) and then managed a PhD right before 40, so I'm hannnnggging onto early career researcher as long as possible, its like the fountain of youth for someone who transitioned into research later in life. ðŸ«
comment in response to
post
📌
comment in response to
post
Not sure if this fully fits but we used zenodo to point to our data and GitHub at the time of publication. zenodo.org/records/7113...
comment in response to
post
I'm quite a fan of this paper, about how the millions we've spent looking into the genetic determinants of longevity have mainly been looking at genetic determinants of pension fraud www.biorxiv.org/content/10.1...
comment in response to
post
I agree, it's very baffling to see a paper where they have carefully sampled across development and then be like here's some pseudo-time analysis. I wrote my PhD thesis on a pseudo time algorithm, and I find it odd.
comment in response to
post
Things get more complicated when you are trying to map dynamic cell states (development or disease), and that's where lower dimensional embeddings can lead you astray. This type of analysis is more likely now and the field is still working it's way through transitioning. Again just my lousy 2 cents.
comment in response to
post
It's a good question. My complete guess is, they were sexy and informative early on. When you run static states, fully differentiated cell states through them the islands of cell types paint an easy to follow narrative. Here are my clusters/cell-types next plot here are what they are expressing.
comment in response to
post
I think you are correct. There is no lower dimension that preserves the ground truth it's mathematically impossible. I treat them as a hypothesis generator for future statistical analysis or as a story teller for the analysis I'm about to show the reader. If that makes sense?
comment in response to
post
The connection the poster implies (in replies) is that neuro-imaging data is quantitative continuous and so is scRNA-seq over cell state changes. I don't believe this is the case. The nonzero variation of transcripts is generally low and publications have shown treating it as binary is beneficial.
comment in response to
post
Tricky! Lower-dimensional embeddings involve trade-offs, emphasizing local neighborhoods (t-SNE) at the cost of longer-range interactions. UMAP attempts a balanced approach, while force-directed layouts give you all the options. 😂 All are fair, but remember, they're visualizations, not quantitative.
comment in response to
post
While I generally agree with the poster and the sentiment around over analyzing UMAPs (etc), in this case the proof for poor PCA performance appeared to be extrapolated from a neuro imagining paper without any substantive proof it carries over to scRNA-seq.
comment in response to
post
The post was very specific to PCA and VAE, which I'm not on board with being grouped together, but my guess is that they would agree that PCA is passing artifacts downstream to tSNE and UMAP, both of which should never be used for nothing more than a qualitative assessment that your data isn't 💩
comment in response to
post
Funny enough, while trying to work out a way to quantify how wonky or isometric a scRNA-seq dataset is, I ended up spending part of my evening diving into this evolutionary biology paper titled "Be careful with your principal components" :)
comment in response to
post
@filbednar.bsky.social it was in reference to single cell RNA-seq
comment in response to
post
You are not wrong. I was being more cheeky than an absolutist.