Wired reported that Hugging Face released Carbon-X, an open-source model for answering questions about DNA in ordinary language, on 22 October. The AI model platform developed it with researchers at Zhongguancun Academy in China and the University of Naples Federico II in Italy.
The team taught an existing language model to read DNA in groups of six letters, then trained it further on 300 billion tokens comprising genetic sequences, English and examples connecting DNA to biological descriptions. Carbon-X can take in roughly 50,000 base pairs at once, enough to examine a section of a genome rather than the whole thing.
Reading a genetic test result could involve asking what a short stretch of DNA might do to gene activity, if Carbon-X’s answers prove accurate.
Lewis Tunstall, a senior research scientist at Hugging Face, said the 14-billion-parameter model was designed to run locally, so sensitive data need not go to a company. He said viruses were excluded from its training sequences to avoid potentially dangerous uses.
Carbon-X’s developers caution that they have not validated its predictions in wet-lab experiments and cannot yet say how accurate they are.
