Google DeepMind Achieves Breakthrough in Protein Design

Google DeepMind published a landmark paper today introducing AlphaProteo 2, a next-generation AI system for de novo protein design that achieves an 89% experimental success rate — nearly double the previous best of 46%. The system can design functional proteins for specified therapeutic targets in hours, a process that traditionally took research teams months or years.

AlphaProteo 2 builds on DeepMind's foundational work with AlphaFold, but goes beyond structure prediction to actually designing novel proteins with desired properties. The system was validated through wet-lab experiments at DeepMind's London biology lab and partner institutions, with 89 out of 100 designed proteins folding correctly and binding to their intended targets.

The implications for drug discovery are profound. DeepMind demonstrated that AlphaProteo 2 could design candidate therapeutic proteins for three different disease targets — including a novel anti-inflammatory compound — in under 48 hours each. Traditional approaches to the same targets had been stalled for over two years.

DeepMind is making AlphaProteo 2 available to academic researchers through a managed API and has established partnerships with Isomorphic Labs (its Alphabet sister company) and two major pharmaceutical companies to bring AI-designed therapeutics into clinical trials. CEO Demis Hassabis called the achievement "the beginning of a new era in biotechnology."

The paper has been accepted for publication in Nature and will be presented at the upcoming ICLR 2026 conference. DeepMind also released a smaller, open-source version of the model for academic use, continuing its commitment to advancing scientific research through open AI tools.