Google DeepMind’s new AI can model DNA, RNA, and ‘all life’s molecules’

Google DeepMind is introducing an improved version of its AI model that predicts not just the structure of proteins, but also the structure of “all life’s molecules.” The work from the new model, AlphaFold 3, will help researchers in medicine, agriculture, materials science, and drug development test potential discoveries.

Previous versions of AlphaFold only predicted the structures of proteins. AlphaFold 3 goes beyond that and can model DNA, RNA, and smaller molecules called ligands, expanding the model’s capability for scientific use.

DeepMind says the new model shows a 50 percent improvement in prediction accuracy compared to its previous models. “With AlphaFold 2, it was a big milestone moment in structural biology and has unlocked all kinds of amazing research,” DeepMind CEO Demis Hassabis told reporters in a briefing. “AlphaFold 3 is a step along the path in terms of using AI to understand and model biology.”

AlphaFold 3 has a library of molecular structures. Researchers input a list of molecules they want to combine, then AlphaFold 3 uses a diffusion method to generate a 3D model of the new structure. Diffusion is the same type of AI system that AI image generators like Stable Diffusion use to assemble photos.

DeepMind says Isomorphic Labs, a drug discovery company founded by Hassabis, has been using AlphaFold 3 for internal projects. So far, the model helped Isomorphic Labs improve its understanding of new disease targets.

Along with the model, DeepMind is also making the research platform AlphaFold Server available to some researchers for free. The server, powered by AlphaFold 3, lets scientists generate biomolecular structure predictions regardless of their access to compute power. Hassabis says the server is available for academic, non-commercial uses, but Isomorphic Labs is working with pharmaceutical partners to use AlphaFold models for drug discovery programs.

Google says it is working with the scientific community and policy leaders to deploy the model responsibly. Google says in a paper that some biosecurity experts believe AI models could “may lower the barrier for threat actors and enable them, in concert with other technologies, to design and engineer pathogens and toxins that are more transmissible or harmful.”

The company says it worked with domain experts and biosecurity, research and industry specialists to figure out risks around AlphaFold 3 even before its launch.