Claude models trained on protein sequences at scale, tests show
Reported by Hacker News and news.lavx.hu2 sources ↓
Tests reported by Hacker News found that Claude Fable 5.2 and Opus 5.5 predicted amino acids in little-studied proteins, while tested GPT, Gemini and open-weight models stayed near zero on the same task.
Key points
- Fable 5.2 predicted the next amino acid correctly 54% of the time on real proteins and 9% on shuffled sequences
- The test covered 668 bacterial and archaeal proteins from 167 families with few papers and no close UniProt relatives
- Fable 5.2 outperformed the open protein language models tested; Opus 5.5 matched ProtGPT3-10B
- The testers said Claude's accuracy rose with the number of sequences in a protein family
Sources
Details
The test used protein sequences from families of unknown function, with no solved structure or more than three papers per family. For each protein, testers prompted models with sequence prefixes and scored their predictions of the next amino acid, comparing results with shuffled versions of the same proteins.
Hacker News reports that Fable 5.2 scored 54% on real sequences and 9% on shuffled ones. Opus 5.5 performed about as well as ProtGPT3-10B, while Fable 5.2 scored higher than the open protein language models included in the comparison. GPT, Gemini and seven tested open-weight models stayed near zero on the task.
In a wider test of 257 protein families, the reported Claude scores increased with the number of sequences per family; Fable 5.2's gains leveled off at about 300 proteins per family. The testers also said the models' sequence knowledge carried over to proteins published after their training cutoff, where there was no close earlier relative. The article does not establish how the protein sequences entered training data.
Written from coverage as of
Images
- Hacker News
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh