Study finds latest Claude models were trained on protein sequences at scale
LeoTZ03 · x · 2026-11-18
Researchers ran a next-amino-acid test on 20 LLMs, giving them the start of little-studied bacterial proteins. Only the latest Claude models (Fable 5.2 and Opus 5.5) predict the next amino acids—not GPT, Gemini, or seven open-weight models.
Method and evidence:
- Test set: 668 bacterial/archaeal proteins from 167 Pfam families of unknown function—no solved structures, ≤3 papers per family, no close UniProt relatives
- When proteins are shuffled (order destroyed, letters kept), accuracy collapses, showing the models know the sequence itself
- Behavior mirrors protein language models: better on larger families, even families with zero papers
Results: Fable 5.2 beats every open protein language model tested; Opus 5.5 is level with ProtGPT3-10B, the largest.
The motivation was to distinguish whether frontier LLMs learn from sequence databases like UniProt or only from text about proteins—famous proteins like GFP can't separate the two, hence the obscure test set.
Related event: Test Suggests Latest Claude Models Trained on Protein Sequences at Scale(3 posts)→
More from Models
- Gemini cracks an unread 1466 Sforza cipher by hunting down a 1465 key in Vienna — fofrAI · 2026-11-18
- Cloudflare's Clairic extraction model returns confidence scores and spans, no preset categories — altryne · 2026-11-18
- Claude's biggest mistakes come from following rules too rigidly — wavefnx · 2026-11-18
- Interact-0: a 31B full-duplex speech model that knows when to speak in groups — niloofar_mire · 2026-11-18
- Asked for "exactly 10 amino acids," Claude hallucinates GPT-7 becoming the 267th Pope — LeoTZ03 · 2026-11-18
- Haider: No Chinese Model Is on the Pareto Frontier Anymore — Compute, Not Algorithms, Is the Bottleneck — haider1 · 2026-11-18