After AlphaFold, DeepMind's AlphaProteo designs brand-new proteins that latch onto disease targets

Google DeepMind introduced AlphaProteo, an artificial-intelligence system that designs entirely new proteins built to bind tightly to chosen target molecules, a capability that could accelerate drug development, diagnostics and biological research. Announced on September 5, 2024, it extends the lab's work beyond AlphaFold, which predicts protein structures, toward actively designing useful new proteins from scratch, one of biology's hardest and most valuable challenges.

The technical results are striking. Given a target protein's structure and the spots where a binder should attach, AlphaProteo generates candidate proteins that, when synthesized and tested in the lab, latched onto their targets with binding affinities 3 to 300 times stronger than the best existing methods, roughly ten times better on average. For one target, the viral protein BHRF1, 88 percent of the AI's candidates bound successfully in wet-lab testing. The system was trained on experimental structures from the Protein Data Bank plus more than 100 million structures predicted by AlphaFold.

“Announced on September 5, 2024, it extends the lab's work beyond AlphaFold, which predicts protein structures, toward actively designing useful new proteins from scratch, one of biology's hardest and most valuable challenges.”

The potential applications span health and biology. AlphaProteo designed successful binders for proteins tied to viral infection, including the SARS-CoV-2 spike protein, and to cancer and inflammation, such as VEGF-A and PD-L1. Strong, custom-made binders are useful as building blocks for new medicines, for biosensors that detect disease, and for imaging cells and tissues, potentially shortening work that today can take many experimental rounds. Researchers at the Francis Crick Institute independently confirmed some of the binding interactions.

The caveats are clear and openly acknowledged. AlphaProteo did not succeed everywhere; it failed to design a working binder for TNF-alpha, a protein linked to rheumatoid arthritis, showing the method still struggles with the hardest targets. Designed proteins also require extensive laboratory and safety testing before any clinical use, and binding tightly is only one of many properties a real therapeutic needs. Even so, an AI that reliably designs novel, high-strength protein binders is a powerful new tool for the scientists working to understand disease and build the medicines of the future.