prion-fasta

This is a collection of FASTA sequences tagged 'prion' on the US National Center for Biotechnology Information (NCBI) website.

Understanding

IUB/IUPAC set a format where the letters A-Z are used to represent a chain of amino acids, instead of spelling out ACGT/U of DNA and RNA. Many lines of the file are accessions of Major prion protein from humans and other mammals. This protein is believed to normally serve a complex role in the brain and nervous system. Our current understanding is that a misfolded form of this protein causes more of this protein to be folded in the same way. This misfold is transmissible to other animals, and the cause of BSE in cows, scrapie in sheep, CWD in deer, a disease in camels, and CJD, vCJD, and kuru in humans.

Curation and Duplication

A recent biology + deep learning paper found just under 3,000 prion proteins on the UniProt database. This file contains 45,977 lines. What's going on?

I did limited filtering. For example proteins from a plant genome coauthored by "Florent Prion" and descriptions containing "virus" were removed.

Proteins such as KAH3171747.1 are from a fungus and labeled "Prion-inhibition and propagation". You can search IDs on https://www.ncbi.nlm.nih.gov if you have additional questions about specifics.

Some context

• 14,920 lines are "hypothetical protein"
• 4,605 are partial
• 2,154 are "sequence \d+ from patent"
• 1,342 are prion-like
• 392 are "Homo sapiens"
• 212 are peptides and
• 198 have the amino acid string "RIKWNRKVLPSTSQVTEARTAEIRPGAFIKQGRKLDIDFGVEGNRYYEANYWQFPDGIHYNGCS"

Prions and Deep Learning Research

On Prions and Protein Design - "Learning to use prions as nature’s laboratory for modeling protein dynamics may move us closer to mastering protein design as a whole."

Deep learning reveals antimicrobial peptides within prions - "screened 19.3 million fragments from 2,897 curated prion-related proteins and identified 1,179 candidate antimicrobial peptides"; concluded "antimicrobial potential is distributed broadly across prion-associated sequence space".

Their PyTorch model is somewhat old; more recent projects do finetuning by SFT or X-LoRA, and use ProGen or ESM2 as the base model.