Georeactor Blog<br>RSS Feed

Folding Prion Protein with AI<br>August 11, 2026

Tags: bioseries<br>Jump to:

AI safety says no

How science discovered prion diseases

Current AI research into folding prion protein

Compiling a dataset

Convincing ChatGPT to give me a lift

Actually folding

Concluding thoughts

Back in the olden days, I remember the SETI @ home screensaver. This quickly inspired Folding @ home, and a pop science article introduced me to the concept of protein folding.

I recently idly made a list of domains which would be interesting to train or finetune a genomic language model (gLM) for: long non-coding RNA, transposable elements, epigenetics, or prions.

I'd read a little on prions four years ago from the book Madness and Memory. While I was researching this post, the idea got more relevant with scrapie detected in one sheep in Oregon.

AI safety says no<br>You might think, ⚠ ️let's pump the brakes on the prion stuff.

When I asked "is there a dataset of prion genomes / sequences somewhere?" Claude refused to discuss it, and ChatGPT ultimately erased its answer, suggesting I could apply to the trusted testers program.

But major prion protein plays a complex role in a healthy nervous system. AlphaFold already has an entry and 3D model in their Protein Structure Database, and I haven't seen anything suggesting that it would be excluded from sequence-based gLMs such as AlphaGenome or Evo (context: Evo excluded eukaryote-infecting viruses from training data. I wondered if it would be effective or like leaving a tricycle out of training data with other-wheeled vehicles (analogy ok?) but the Evo 2 paper seemed to back it up).

Even after all this reading and clicking and stuff, I'd never folded any kind of protein before, so I'd like to see what I can learn. Thanks to Qwen for making this cute infographic.

How science discovered prion diseases<br>Major prion protein has a common ancestor 100 million years ago, "with placental mammals possibly being generally susceptible to disease".

The first records of scrapie (TSE prion disease in sheep) appear ~300 years ago in the sheep of East Midlands in England.

In the 20th century, this class of diseases became known as "slow viruses".

In 1982, Stanley Prusiner traced scrapie to a prion protein by eliminating any other possible agent; he would later win the Nobel Prize. When the UK identified an outbreak of BSE in cows later in the 80s, interest intensified in understanding the cause.

A 1991 paper from Jean-Pierre Liautard "Are prions misfolded molecular chaperones?" describes a circulating hypothesis, which is maybe the first example in literature to describe prions as "misfolded". Prusiner was on a similar track at the time, authoring or co-authoring papers finding "two different topological forms in cell-free systems" and "an abnormal isoform of the prion protein, which is encoded by a chromosomal gene". There was maybe a hesitation around when and how prion protein acquired an unusual shape; whether it was during folding at the ribosome, alternative splicing, or an additional component attached to the protein.

I've tried to figure out when scientists learned about protein folding and misfolding generally, and it must go back a long way (Francis Crick wrote about enzymes and "misfolding of the protein" as early as 1964).

Around 1994 the prion protein's structure had been simulated on computers, and an actual X-ray image was captured in 2001.

Current AI research into folding prion protein<br>The 2025 paper Diffusing protein binders to intrinsically disordered proteins used the open source protein-folding model RFdiffusion to create a binder for prion protein, which could be useful for imaging.

Eryney Marrogi and Theodore Sternlieb posted an article about "Learning to use prions as nature's laboratory for modeling protein dynamics may move us closer to mastering protein design as a whole."

They write that AlphaFold's outputs "excel at predicting a protein's most stable form, [but] they fail to predict how proteins 'wiggle' and 'shake' between multiple structures". Prions are an extreme example, but they suggest studying them in a safe, virtual environment.

Then I saw this paper in June: Deep learning reveals antimicrobial peptides within prions. These researchers had taken 2,897 known prion sequences, and sent fragments through their peptide antibiotic classifier (a PyTorch model named APEX). Apparently this connection of prions to the immune system has been circulating for years (see: The antimicrobial protection hypothesis of Alzheimer's disease, 2018).

This AI-accelerated search led the researchers to conclude that "encrypted antimicrobial potential is distributed broadly across prion-associated sequence space". This is yet another reason to learn more about prion proteins.

Compiling a dataset<br>You're probably familiar with 'mad cow disease'. In sheep this is called scrapie, in cows it was BSE, and in humans it's CJD or vCJD; but ultimately these are all...