Pinal designs proteins from natural-language function prompts
Original title: Vibe genes? Natural(ish) language comes to protein, genome design software
A China-based team tested a fluorescent protein, two enzymes, and a complex subunit; Hugging Face hosts a DNA tool.
Researchers based in China built Pinal so users can describe a protein function in plain language and receive an amino-acid sequence. T2struct turns the prompt into spatial tokens, and SAProt-T fills in the sequence after a global recheck. Training used function tags expanded by an LLM, sequences from three databases, and AlphaFold-predicted structures, including one set of 800 million keyword–protein pairs. The team asked for a fluorescent protein, two enzymes, and part of a multi-protein complex. Hugging Face separately said it is hosting a similar DNA-design tool.
Unconfirmed: The excerpt ends before stating how many of the four test designs actually worked.
- T2struct outputs neighbor-relationship tokens, not a 3D shape.
- SAProt-T re-evaluates those probabilities against the original prompt.
- One training database held 800 million keyword–protein pairs.
- AlphaFold errors may affect designs of obscure proteins.
- The article says a working protein is often not the result.
- Organizations
- Hugging Face
- Countries
- China
From the original article
8 min read at arstechnica.com
You can apparently now tell AI tools what sort of protein you want them to make.
By John Timmer · arstechnica.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.