Claude models trained on protein sequences at scale, study finds
New tests show Claude Fable 5.2 and Opus 5.5 predict amino acids in obscure proteins, indicating large-scale training on biological sequence data.
この記事は英語版のみ利用可能です。
A technical investigation published in November 2026 reveals that the latest Claude models were trained on protein sequence data at a significant scale. Researchers Xinming Tu and Leo Tianlai Chen demonstrated that Claude Fable 5.2 and Claude Opus 5.5 can accurately predict the next amino acid in little-studied bacterial proteins, a capability absent in other frontier models. This finding suggests that Anthropic included vast biological databases in its training corpus, distinguishing its approach from competitors like OpenAI and Google.
What happened
The study focused on next-amino-acid prediction using a test set of 668 bacterial and archaeal proteins from 167 Pfam families of unknown function. These proteins were selected because they have no solved structures, fewer than three associated papers, and no close relatives in standard databases like UniProt. By cutting each protein at five points and asking the model to predict the next ten amino acids, the researchers measured how well the models understood the underlying biological sequence rather than just reciting memorized text from scientific literature.
The results showed a stark contrast between Claude models and other leading systems. Claude Fable 5.2 achieved 54% accuracy on real protein sequences compared to only 9% on shuffled versions where the order was destroyed. Claude Opus 5.5 showed similar behavior, with accuracy rising significantly from earlier versions. In comparison, GPT-6.2 Sol, GPT-6 Astra, Gemini, and seven open-weight models including DeepSeek V4.1 Flash and Qwen 3.8 remained near zero accuracy, even when prompted to continue raw FASTA text. This indicates that while other models may know about famous proteins like GFP from reading papers, they have not internalized the statistical structure of protein families at scale.
Further testing confirmed that this capability stems from training data rather than reasoning or web search. The models performed equally well on protein families with zero published papers, ruling out the possibility that they were learning from textual descriptions. Additionally, the accuracy improved with larger protein families, mirroring the behavior of specialized protein language models. The knowledge also extended to 56 proteins made public in August 2026, after the reported training cutoff, suggesting the models can generalize to new sequences based on learned family structures.
How it works
The core mechanism tested is next-token prediction applied to biological sequences. Protein language models typically read a sequence of amino acids from left to right and predict the most likely next residue based on patterns learned from millions of other sequences. The researchers applied this same logic to general-purpose large language models. By providing the start of a protein sequence and asking for the continuation, they tested whether the model had learned the joint probability distribution of amino acids within specific families.
To ensure the models were not simply guessing based on amino acid composition, the researchers used a control method involving shuffled sequences. Shuffling keeps the same letters but destroys the biological order. A model that has truly learned protein structure will perform well on real sequences but poorly on shuffled ones. The gap between these two scores measures the model's understanding of sequence order. Claude Fable 5.2 and Opus 5.5 showed large positive gaps, while other models did not, indicating that Claude’s training data included enough protein sequences to learn these complex dependencies.
Key details
- Claude Fable 5.2 achieves 54% accuracy on next-amino-acid prediction for obscure proteins, beating open protein language models like ProGen3-3B.
- Claude Opus 5.5 performs at a level comparable to ProtGPT3-10B, one of the largest open protein language models available.
- Other frontier models, including GPT-6.2 Sol, GPT-6 Astra, and Gemini, show near-zero accuracy on the same tasks, indicating they were not trained on protein sequences at scale.
- The models perform well on protein families with no published papers, proving the knowledge comes from sequence data rather than scientific literature.
- Accuracy correlates with family size, with performance leveling off at around 300 proteins per family, consistent with how protein language models learn.
- The capability extends to proteins published after the training cutoff, demonstrating generalization to new sequences within known families.
Why it matters
For software engineers and AI researchers, this finding highlights a divergence in training strategies among frontier model providers. While some companies focus primarily on text and code, Anthropic appears to have integrated large-scale biological data into its general-purpose models. This suggests that future LLMs may increasingly serve as multi-modal tools capable of handling specialized scientific domains without needing external plugins or fine-tuning. It raises questions about data transparency and the potential for unexpected capabilities in general-purpose models.
From a practical standpoint, this means developers building bioinformatics tools might be able to leverage general-purpose Claude models for tasks previously reserved for specialized protein language models. If Claude can predict sequences and potentially rank variants or infer structure, it could simplify pipelines that currently require multiple specialized models. However, it also implies that users should be aware of the potential for these models to contain sensitive or proprietary biological data if such data was present in the training sets.
What you can do
- Test Claude models on your own obscure protein sequences to see if they provide useful predictions for variant ranking or structure inference.
- Compare the output of Claude Fable 5.2 against specialized tools like ESMFold or AlphaFold for quick preliminary analysis.
- Be cautious when using general-purpose LLMs for sensitive biological data, as they may have memorized sequences from their training set.
- Monitor updates from Anthropic regarding data sources and training methodologies to better understand the scope of their biological knowledge.
- Consider using Claude for initial hypothesis generation in protein engineering tasks, leveraging its ability to generalize across protein families.
- Review your data privacy policies if you are uploading proprietary protein sequences to any cloud-based LLM service.



