Recently I collected ~46,000 sequences labeled 'prion' on the NCBI website. In this post I use gLM2 embeddings to try and group the dataset into meaningful clusters.

In my initial post I noticed ~200 sequences have the same amino acid sequence, ~400 are human, and the dataset includes proteins which are prion-like or inhibit prions... just anything which was associated with prions. I'd like to see clusters like that.
Right away the script removed 7 sequences which were a little longer than the model allowed, but nothing as bad as the complete genomes which came through before my filtering in my previous work. I decided to review if those belong in the dataset. A rat protein was linked to a paper with "prion" in the title, but I couldn't tell how relevant this specific protein was. There was one long sequence (XP_072180894.1) where I can't figure out what the prion connection might be? I would rate this sample as 'OK-ish'.

The clustering diagram essentially confirms there are a few large groups, and most are disconnected:

I asked ChatGPT for some code to pull descriptions of proteins in the most popular clusters, and then manually investigated any commonality:

1, Most popular: prion inhibition
2, unusual variety?
3, "prion-like"
4, "shadow of prion protein"
5, "prion formation protein"
6, "domain-containing protein"
7, proteins from a wheat fungus Parastagonospora nodorum
8, prion-inhibition
9, "shadow of prion protein 2"
10, human prion protein

Somewhat related - there is a recent article that frequency of a specific genetic sequence / mutation rate is a bad measure for importance of that sequence: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3004089

Georeactor Blog

Clustering Prion Proteins