Pinal designs proteins from text: one increased product yield by a factor of 1.7
Biotech needs a faster way to select molecules for laboratory testing. Pinal takes a description of a protein's function and proposes its structure and amino acid sequence: one resulting candidate increased product yield by a factor of 1.7 in an experiment. But finding a fluorescent protein required generating around 50,000 sequences. Meanwhile, Carbon-X has been released as an open model for questions about DNA.
AI-processed from Habr AI; edited by Hamidun News
A team searching for a new enzyme needs a list of molecules worth testing in the laboratory. Pinal allows that search to start with a text description of the desired function. In one experiment, a protein it produced increased product yield in an artificial CO₂ fixation pathway by a factor of approximately 1.7 compared with its natural counterpart from E. coli. For biotech developers, that is a reason to test whether the model can help select candidates ahead of expensive experiments.
Another recent arrival is Carbon-X, an open model for analyzing DNA through ordinary conversation. The systems address different tasks: Pinal proposes proteins, while Carbon-X helps interpret genetic sequences. Their shared product concept is straightforward: a specialist asks a question in words, and the system connects it to biological data.
Pinal: from a function description to laboratory candidates
Imagine you need an enzyme that performs a particular reaction. The user describes the function, and Pinal then works in two stages. The T2struct module proposes a structural representation of the protein. SaProt-T then uses that representation, together with the text prompt, to select an amino acid sequence: the chain that makes up the protein.
According to the authors, the system was trained on 1.7 billion pairs of protein data and text descriptions, with synthetically generated annotations. This is a significant caveat: the size of the training set should not be interpreted as the number of independently verified laboratory observations. The authors describe the architecture and training approach in their Nature paper.
In the experiments presented, the researchers found working molecules for four tasks. However, the fluorescent protein experiment offers the most revealing result for budget planning. The system generated around 50,000 sequences, screening narrowed the pool to 29 candidates, and one actually proved fluorescent.
These are distinct stages of selection, so collapsing them into a single universal measure of “model accuracy” would be a mistake. They do, however, show the actual process: generating a large pool, computational filtering, and then a physical experiment. For a product manager, the cost of finding a working candidate matters more than how polished the answer looks in the interface.
Carbon-X: ask a question about DNA in a chat
Carbon-X is built on Qwen3-14B and has 14.8 billion parameters. It was released by researchers from Hugging Face, Zhongguancun Academy, and TIGEM / the University of Naples. Its code and weights are available under Apache 2.0. The model accepts DNA sequences alongside questions in natural language; the model family and terms of use are described in the model card.
One simple use case: a researcher supplies an original DNA segment and a variant with a single-letter substitution, then asks about the possible consequences. Another is to ask whether a fragment resembles a region that regulates gene activity. Such an interface could help a team building tools for researchers reduce manual switching between sequences, annotations, and explanations.
But a coherent explanation does not establish that a biological conclusion is correct. The developers warn of errors in base counting, coordinates, and precise sequence changes. When context is insufficient, the model can invent a function for a region. The quality of long DNA sequence generation also declines.
What to check before integration
My practical takeaway is to start with a narrowly defined operation for which verified examples and a clear understanding of the cost of errors already exist. For Carbon-X, that could mean preliminary analysis of known variants followed by expert review. For Pinal, it could mean proposing candidates within an existing laboratory workflow.
Precise sequence operations are best assigned to code: checking length, locating a substitution, and comparing the original and modified fragments. The model can handle explanations and hypothesis generation. In the interface, retaining the input data, model version, and external verification results helps make answers reproducible.
For biotech founders, these projects provide a concrete reason to run a pilot. It should be assessed by the number of validated results, specialist time, and experimental costs. A successful individual protein or a convenient DNA chat interface alone does not establish how much solving your particular problem will cost.
- Nov 12, 2026AI agents re-checked 2,000+ ICML 2026 papers and found errors that reviewers missed
- Nov 12, 2026Eight Countries, OpenAI and Anthropic Promise an AI "Kill Switch" — but Set No Deadlines or Penalties
- Nov 12, 2026Hugging Face warns: a wave of hacks could make open-source AI less open
- Nov 12, 2026OpenAI Built a Control Room: A Kill Switch for Models That Escaped the Sandbox
Key Questions Answered
What are the key details on Pinal: from a function description to laboratory candidates?
Imagine you need an enzyme that performs a particular reaction. The user describes the function, and Pinal then works in two stages. The T2struct module proposes a structural representation of the protein. SaProt-T then uses that representation, together with the text prompt, to select an amino acid sequence: the chain that makes up the protein. According to the authors, the system was trained on 1.7 billion pairs of protein data and text descriptions, with synthetically generated annotations.
What are the key details on Carbon-X: ask a question about DNA in a chat?
Carbon-X is built on Qwen3-14B and has 14.8 billion parameters. It was released by researchers from Hugging Face, Zhongguancun Academy, and TIGEM / the University of Naples. Its code and weights are available under Apache 2.0. The model accepts DNA sequences alongside questions in natural language; the model family and terms of use are described in the model card.
What to check before integration?
My practical takeaway is to start with a narrowly defined operation for which verified examples and a clear understanding of the cost of errors already exist. For Carbon-X, that could mean preliminary analysis of known variants followed by expert review. For Pinal, it could mean proposing candidates within an existing laboratory workflow. Precise sequence operations are best assigned to code: checking length, locating a substitution, and comparing the original and modified fragments.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.