Anthropic's Claude Opus 5.5 and Fable 5.2 show signs of protein-sequence training
A third-party test found that Claude Opus 5.5 and Claude Fable 5.2 can predict the next amino acid in little-studied bacterial and archaeal proteins, but fail when the same sequences are shuffled, indicating they learned real sequence patterns rather than just text about proteins. The researchers report no other frontier model tested showed this ability, and Claude's performance rivaled or matched dedicated open protein language models like ProtGPT3-10B.
GoKawiil's interpretation of the reporting above, not reported fact.
The findings suggest Anthropic may have trained its latest models on large protein-sequence databases such as UniProt, a capability not previously confirmed for general-purpose language models. If accurate, this could position Claude as a tool for biological research tasks like variant scoring, though the researchers' test used only 668 proteins from 167 families, leaving open questions about scale and robustness of this capability.
- Claude Opus 5.5 and Fable 5.2 predict next amino acids in little-studied proteins but fail when sequences are shuffled
- No other frontier model tested showed this sequence-specific behavior
- Claude's performance on this task matched or exceeded dedicated open protein language models like ProtGPT3-10B
Source: xinmingtu.cn — Xinming Tu, 2026-11-18
Published there as: “Claude was trained on protein sequences at scale”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.