Hugging Face's Carbon-X lets researchers chat with DNA

Hugging Face developed Carbon-X with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy, expanding an existing language model's vocabulary to read DNA in groups of six letters. The training data combined DNA sequences from mammals including humans, other vertebrates, invertebrates, plants, fungi, and protozoa with English text and examples linking DNA to biological descriptions.

The model enters a field where Google DeepMind's AlphaGenome predicts the effects of DNA sequences and mutations in a numeric format, and Evo 2, developed by the Arc Institute and Nvidia with academic partners, can generate entire genomes from scratch. Carbon-X instead integrates DNA and natural language in a conversational interface, which Tunstall describes as unlocking the ability to chat with DNA.

The team intentionally kept the model small at 14 billion parameters so it can run locally, avoiding restrictions researchers face when uploading identifiable or sensitive patient data to frontier AI models. Hugging Face, which Nvidia agreed in September to acquire for nearly $13 billion, positions Carbon-X for drug and genetic therapy design, but the developers caution they have not validated it with wet lab experiments, so accuracy claims remain unproven.