Blog finds latest Claude models predict obscure protein sequences
Original title: Claude was trained on protein sequences at scale
Fable 5.2 beat open protein models on 668 little-studied proteins; GPT, Gemini and open-weight LLMs stayed near chance.
A researcher tested whether frontier models can continue little-studied bacterial and archaeal proteins, a task famous proteins cannot separate from memorized papers. Claude Fable 5.2 and Opus 5.5 predicted the next amino acid far above shuffled controls; 20 other language models did not. Fable 5.2 also beat ProGen3-3B and ProtGPT3-10B on the same left-to-right task. An update says the models still predicted 56 proteins first published in August 2026.
Unconfirmed: The author infers sequence-database training; the text does not include confirmation from Anthropic.
- The test set has 668 proteins from 167 unknown-function Pfam families.
- Fable 5.2 scored 54% on real sequences and 9% on shuffled ones.
- Opus 5.5 matched ProtGPT3-10B and trailed ProGen3-3B.
- Accuracy rose with larger families, including families with no papers.
- GPT-6.2, Gemini and seven open-weight models stayed near zero.
- Organizations
- Anthropic
From the original article
15 min read at xinmingtu.cn
The latest Claude models were trained on protein sequences, at scale. Give Claude Fable 5. 2 or Claude Opus 5.
By gmays · xinmingtu.cn
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.