Claude models demonstrate large-scale protein sequence training
New research shows Claude models can predict amino acids in obscure proteins, suggesting they were trained on massive biological datasets.

The short version
- Claude Fable 5.2 correctly predicted the next amino acid in 54% of cases for obscure bacterial proteins.
- Other frontier models like GPT and Gemini showed nearly zero accuracy on these obscure sequences.
- Claude models outperformed open-weight protein language models on many little-studied protein families.
- The models' ability to predict sequences for unknown proteins suggests they learned from vast amounts of raw sequence data.
Why it matters
This discovery reveals that frontier AI models are beginning to internalize complex biological patterns beyond just human language. It suggests that large-scale biological data is becoming a standard component of general-purpose AI training.
Hidden biological knowledge
Researchers have discovered that the latest Claude models possess a deep understanding of protein sequences. By testing the models on obscure bacterial and archaeal proteins, the team found that Claude can predict the next amino acid in a sequence with significant accuracy. This ability is notable because many of the proteins used in the study have no known structure and very few research papers associated with them.
The study used a test set of 668 proteins from 167 families of unknown function. To ensure the models weren't just memorizing well-known data, the researchers chose proteins that have no close relatives in the UniProt database. The results showed that Claude models could handle these
Testing methodology and constraints
To rigorously evaluate the models, researchers performed a total of 3,340 cuts across the protein sequences. Each cut was placed at five random positions within the Pfam domain, which is the longest span of a protein's functional area. These domains were required to be at least 40 residues long, and the cuts were positioned at least 10 residues from either end of the domain with at least 30 residues preceding the cut.
This specific cutting method ensured that the models were being tested on meaningful segments of the proteins rather than just the very beginning or end of a sequence. By providing a prefix and asking for the next 10 amino acids, the researchers could measure how well the models understood the underlying biological logic. This setup allowed for a standardized way to compare different models across a wide variety of different protein families.
Performance of other frontier models
The study highlighted significant limitations in other major frontier models when faced with these biological tasks. For instance, GPT-6.1 Sol was unable to reach a length of 100 residues in its responses, while GPT-5.5 could not get past 2 residues. These results suggest that while these models may be capable in other areas, they lack the specific sequence knowledge that the Claude models demonstrated during this test.
Furthermore, several models like GPT-4o, DeepSeek, Qwen, Kimi, GLM, and gpt-oss initially declined to provide any guesses for the protein sequences. To account for this, the researchers provided these models with a modified prompt that explicitly instructed them to give their best guess even if they were uncertain. Despite these instructions, the performance of these models remained significantly lower than that of the Claude models on the obscure protein sets.
Safety filters and technical nuances
The research also documented how different companies handle biological prompts through their safety layers. OpenAI rejected a variety of requests from models like GPT-6 and GPT-6.1 Sol, citing concerns over possible biological misuse. Specifically, 23 requests were rejected in the first round, and 17 were rejected during the main run, while 6.6% of GPT-6.2 Astra’s questions were blocked. This indicates that many frontier providers have active filters to prevent the generation of potentially sensitive biological information.
On a technical level, the study compared how different protein language models process data. While ProGen3 and ProtGPT3 use a one-to-one mapping where each token represents a single residue, ProtGPT2 uses multi-residue tokens. In the case of ProtGPT2, the probability of a specific residue is calculated as the sum of the probabilities of all tokens that begin with that residue. Understanding these architectural differences helps clarify how various models interpret and predict the building blocks of life.
Statistical rigor and data handling
To ensure the findings were statistically sound, the researchers employed several rigorous methods. They used 95% bootstraps over families with 10,000 resamples to determine intervals. The main test was conducted as a one-sided sign-flip permutation test over families, while the Spearman correlation was used to measure the trend between the family-size gap and the logarithm of the family size.
During the evaluation, the models were run through their own APIs with the temperature set to 0. For the Opus 5.5 and Fable models, the researchers kept the thinking setting at its lowest possible level. While the models might generate internal reasoning, the study only scored the visible reply and did not keep the thinking text itself.
Sources and links
- Codegithub.com/XinmingTu/LLM-bioseq-probe github.com
- Reportingxinmingtu.cn: the article we sourced
First reported by xinmingtu.cn. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.