Hugging Face’s New AI Biology Model Lets You Chat With DNA

Hugging Face has unveiled a new AI model called Carbon-X that allows users to interact with DNA through natural language. The model, developed in collaboration with researchers from China and Italy, is trained on both genomic sequences and natural language, enabling users to ask questions about DNA and receive conversational answers.

Senior research scientist Lewis Tunstall explains that Carbon-X aims to assist researchers in designing new drugs and genetic therapies, as well as help patients interpret their own genomes. The model differs from previous AI models like AlphaGenome, which predicts the effects of DNA sequences in numeric format, or Evo 2, which generates entire genomes from scratch.

Carbon-X integrates both DNA and natural language, providing a conversational interface akin to ChatGPT for DNA. To train the model, the team expanded an existing language model's vocabulary to read DNA in groups of six letters and continued training on 300 billion tokens that combined DNA sequences, English text, and examples linking DNA to biological descriptions.

The DNA sequences used for training came from various organisms, including mammals, vertebrates, invertebrates, plants, fungi, and protozoa. However, the model was not trained on viruses to prevent potentially dangerous applications. Users cannot upload full human genomes, as the model is trained to handle DNA sequences of around 50,000 base pairs.

Researchers can use Carbon-X to determine whether a genetic mutation is benign or disease-causing, aiding in the development of targeted drugs or genetic therapies. To address privacy concerns, the developers designed a small model with 14 billion parameters, allowing it to run locally on users' hardware without sending any data to Hugging Face or other companies.

The developers emphasize that they have not yet validated the model's accuracy with actual wet lab experiments.