Hugging Face has released Carbon-X, an open model that can run locally on a laptop and lets users ask questions about DNA in natural language. Developed with researchers from China’s Zhongguancun Academy and Italy’s University of Naples Federico II, the model combines genome sequences with English text to explore how genetic changes may affect gene activity and health.
Lewis Tunstall, a senior researcher at Hugging Face, said Carbon-X is intended to help researchers design drugs and gene therapies and could eventually help patients interpret their own genomes.
“We’ve unlocked a new feature in the language model, which is basically being able to chat with DNA,” Tunstall said of the new model, Carbon X. Natural-language interaction means that nontechnical users can “actually ask questions about DNA that are not accessible with other tools,” Tunstall said.
Connecting DNA and language
Often described as the GitHub for AI models, Hugging Face provides a library of ready-made tools and a platform for running primarily open-weight and open-source AI applications. Its users include developers, researchers, businesses and people who want to experiment with AI without building systems from scratch. In September, NVIDIA agreed to buy the company for nearly $13 billion. Hugging Face was also at the center of a major OpenAI model hack earlier this year.
DNA is more like computer code than natural language. Rather than using 0s and 1s, it encodes information with the letters A, C, G and T. The human genome contains 6 billion of these letters. Although scientists have sequenced the human genome from end to end, they still do not understand the function of much of it.
AI models are now helping researchers investigate some of those unknowns. Last year, Google DeepMind released AlphaGenome, which predicts the effects of DNA sequences and genetic mutations in numerical form, but it requires more computing power. Evo 2, developed by the Arc Institute, NVIDIA and academic partners, works with DNA sequences and can generate entire genomes from scratch.
Carbon-X combines DNA and natural language in a conversational interface—essentially, ChatGPT for DNA.
How Carbon-X was trained
The team began with an existing language model and expanded its vocabulary so it could read DNA in groups of six letters. It then trained the model on 300 billion pieces of data, or tokens, combining DNA sequences, English text and examples that connect DNA with biological descriptions.
The DNA sequences came from mammals, including humans, as well as other vertebrates, invertebrates, plants, fungi and protozoa. Tunstall said the team excluded viruses because it did not want the model to be used for potentially dangerous applications.
Users cannot upload complete human genomes. Carbon-X is designed to process approximately 50,000 base pairs of DNA at a time, allowing queries about only a small section of the genome.
Potential uses and limitations
Scientists could use Carbon-X to help assess whether genetic mutations are benign or disease-causing. That information is an important part of developing targeted drugs and gene therapies. An intervention aimed at the wrong mutation may have no effect or could cause unintended harm.
Researchers face restrictions on uploading identifiable or sensitive patient data to frontier AI models. To address that concern, Tunstall said the team deliberately built Carbon-X as a smaller model with 14 billion parameters. It can run on a user’s own hardware rather than depending on a cloud service, and Tunstall said no data is sent to Hugging Face or any other company.
Carbon-X’s developers caution that the model has not yet been validated in actual wet-lab experiments. As a result, they cannot yet say how accurate its predictions are.
Source: www.wired.com