A new AI model can connect DNA to human language, enabling scientists to interact with genetic code in new ways. Hugging Face announced the Carbon-X model on Thursday as the latest and most powerful in the company’s Carbon family, a group of AI models built to analyze DNA.
“Carbon-X allows researchers to ‘chat with DNA’ for the first time: They can provide a sequence and ask questions about it in natural language,” Hugging Face research scientist Lewis Tunstall told CNET via email.
The word “chat” doesn’t mean talking to your genome but rather using an AI as a conversational interface, allowing users to paste in a DNA sequence, ask a question in plain English and receive back an interpretation, prediction or explanation — like a chatbot for DNA analysis.
The most popular chatbots, such as ChatGPT, Gemini and Claude, are cloud-based models that don’t run on your computer. The processing happens remotely, and you get the results back. But if you’re working on confidential or proprietary research that you haven’t yet published, you might not want to share it with an AI company. “Researchers need models they can study, adapt and run with sensitive or unpublished sequences on their own hardware,” said Tunstall.
One of the nice things about smaller, specialized AI models like Carbon-X is that it can run without using someone else’s computer. Hugging Face says the model can also work in tandem with other AI agents.
DNAgents
To be clear, Carbon-X is not a lab machine, so it can’t physically sequence DNA. It can look at DNA, such as a DNA sequence in a text file, and you can ask questions about it. Tunstall said Carbon-X’s strongest area is the model’s ability to predict mutation effects and complete DNA sequences.
“Carbon-X is particularly strong at helping researchers investigate genetic variation, for example, predicting whether a change in a DNA sequence is likely to be harmful,” Tunstall said.
He also said that it can give another AI model access to a DNA specialist when it needs to extract information.
However, the model does have limitations, according to the company’s blog post. Carbon-X can mess up sequencing operations through miscounts, and it can also hallucinate, which means the model just makes up information. Hugging Face noted that the model may invent the identity or function of a sequence when a user does not provide sufficient information in the prompt.
The quality of Carbon-X’s output also tends to degrade when users ask the model to produce particularly long DNA sequences, according to the company.
Lightning fast
In September, Nvidia announced it would buy Hugging Face for a little under $13 billion. At the time, Nvidia CEO Jensen Huang called the company a “vibrant home for the open model developer community.” Open models are AI models in which some aspects of how they were trained are publicly available.
The release of Carbon-X comes just weeks after OpenAI models began solving longstanding mathematics problems. The AI lab drew widespread backlash from academic researchers after its proprietary models solved several long-standing problems without consulting scientists or showing what work went into solving them.
But Carbon-X is an open model, so AI-savvy biologists can at least peek under the model’s proverbial hood and glimpse how it runs.
“We think researchers also need open alternatives whose capabilities, limitations and risks they can examine independently,” Tunstall said. “By releasing Carbon-X and its training recipe openly, we enable the wider scientific community to test where it works, identify where it fails and help improve it.”
Read the full article here
