Driven on 300 billion tokens mixing DNA and text, the model responds in natural language to the effect of genetic variations and rotates locally, without sending data. It has not yet been validated in the laboratory.
Hugging Face unveiled Thursday Carbon-X, an open weight model developed with the Academy of Zhongguancun, China, and Federico II University of Naples. It accepts both genetic sequences and questions written in common language, and answers in a conversational way about the influence of DNA modification on gene activity or health. Where Google's AlphaGenome DeepMind delivers numerical predictions and where Evo 2 generates genomes, Carbon-X aims for non-specialist use.
How it was built
The team took an existing language model, expanded its vocabulary to read DNA in blocks of six bases, then continued training on 300 billion tokens combining sequences, English text and examples combining DNA and biological descriptions. Genomes come from mammals, other vertebrates, invertebrates, plants, fungi and protozoa; viruses have been deliberately discarded to limit dangerous uses. The model deals with fragments of about 50,000 base pairs, not whole genomes.
Its modest size (14 billion parameters) is a choice: it works on a laptop, allowing researchers to keep data from sensitive patients at home. It could help distinguish a benign mutation from a pathogenic mutation before designing a targeted treatment, but its authors warn that no laboratory experiment has yet measured the reliability of its predictions.
Read the full article on the original source
Go to the source →