Biological language model

A language model trained on DNA, RNA or protein sequences, learning the statistical grammar of biology the way chatbots learn text.

• Trained to predict hidden or next letters across hundreds of millions of natural sequences, it learns which variations evolution tolerates.
• That lets it score mutations: changes the model finds 'surprising' are more often harmful, which helps diagnose rare diseases.
• Generative versions propose new enzymes or antibodies; the Arc Institute's Evo 2 learned from about 9 trillion DNA letters.
• Natural-looking sequences can still fail to fold or work, so every design needs lab testing, and that testing is the bottleneck.