█

AI/TLDR

Xinming Tu · 2026-11-17 · notable

Claude knows proteins — Fable 5.2 and Opus 5.5 predict real amino acid order

In short

Xinming Tu's probe finds Claude Fable 5.2 and Claude Opus 5.5 predict the next amino acid in obscure proteins far better than chance. GPT, Gemini and seven open models show no such skill. Code and data are open.

Xinming Tu · 2026-11-17 · 3 sources checked

Social card for Xinming Tu's study of Claude models predicting protein sequences

A simple next-amino-acid test suggests the newest Claude models were trained on protein sequences at scale.

What is it?

Xinming Tu's study argues that Claude Fable 5.2 and Claude Opus 5.5 appear to have been trained on protein sequences at scale. On 668 little-studied bacterial and archaeal proteins from 167 Pfam 'domain of unknown function' families, Fable 5.2 names the next residue correctly 54% of the time on real sequences.

How does it work?

The probe cuts each protein at five points (3,340 cuts) and scores the first residue a model predicts, at temperature 0 with no tools. A shuffled copy of every protein acts as a control: accuracy on shuffled sequences reflects amino acid frequency only, so the real-minus-shuffled gap measures knowledge of sequence order. Fable 5.2 gains +45.1 points and Opus 5.5 +27.4; GPT, Gemini and seven open-weight models stay near zero.

Why does it matter?

By this measure Fable 5.2 beats the protein language models ProGen3-3B and ProtGPT3-10B, and Opus 5.5 is level with ProtGPT3-10B. The gap also holds in 128 families with no papers and on 56 proteins first published in August 2026, which the author reads as learned sequence knowledge rather than memorized literature. The post notes a limit: the post-cutoff result assumes a 1 July training cutoff.

Who is it for?

Computational biologists and LLM evaluation researchers

Try it

git clone https://github.com/XinmingTu/LLM-bioseq-probe

Sources · 3 outlets

Tags

  • article
  • claude-fable-5-2
  • claude-opus-5-5
  • anthropic
  • protein-sequences
  • training-data
  • evaluation
  • proteomics
  • computational-biology
  • protein-design
  • bioinformatics

← All releases · Learn AI