Hugging Face Releases Carbon-X, a Laptop-Ready AI Model You Can Chat With DNA

Hugging Face has released a new open-source model for understanding DNA that can run locally on a laptop. The model, called Carbon-X, was developed with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy. It is trained on genomic sequences and natural language, allowing users to ask questions about how DNA changes affect gene activity and health and receive conversational answers.

Lewis Tunstall, a senior research scientist at Hugging Face, said the tool is intended to help researchers design new drugs and genetic therapies, and could potentially help patients interpret their own genomes. "We've unlocked a new capability in language models, which is fundamentally about being able to chat with DNA," he said, adding that natural language lets even nontechnical users ask questions about DNA that other tools wouldn't make accessible.

Hugging Face, which provides a library of readymade AI tools and a platform for running open-weight and open-source AI applications, agreed in September to be acquired by Nvidia for nearly $13 billion.

DNA resembles computer code more than natural language, using As, Cs, Gs and Ts instead of 0s and 1s; a human genome contains 6 billion of these letters, and scientists still don't understand how most of it works. Google DeepMind released AlphaGenome last year, which predicts the effects of DNA sequences and mutations in numeric format and requires more compute power, while Evo 2, developed by the Arc Institute and Nvidia with academic partners, can generate entire genomes from scratch. Carbon-X combines DNA and natural language in a conversational interface-like a ChatGPT for DNA.

The team started with an existing language model and expanded its vocabulary to read DNA in groups of six letters, then continued training on 300 billion tokens combining DNA sequences, English text and examples linking DNA to biological descriptions. The sequences came from mammals including humans, other vertebrates, invertebrates, plants, fungi and protozoa. Tunstall said the team did not train the model on viruses to avoid potentially dangerous applications.

Users cannot upload full human genomes; the model was trained to ingest sequences of roughly 50,000 base pairs, so only small parts of the genome can be queried at a time. One use case is helping determine whether a genetic mutation is benign or disease-causing, a key step in developing targeted drugs or genetic therapies. Because researchers face restrictions on uploading identifiable or sensitive patient data to frontier AI models, the team intentionally built a small model with 14 billion parameters that runs on the user's hardware rather than in the cloud, with no data sent to Hugging Face or any other company, Tunstall said.

The developers caution that they have not yet validated the model with actual wet lab experiments, so they cannot say how accurate its predictions are.

Source: Wired science