DeepMind's AlphaProteo designs novel protein binders

Google DeepMind introduced AlphaProteo, a system for designing novel proteins that bind tightly to a specified target molecule — a task central to developing new drugs, biosensors and diagnostics. The model was trained on structures from the Protein Data Bank together with more than 100 million structures predicted by AlphaFold, DeepMind’s earlier protein-structure prediction system, and generates candidate “binder” proteins for a given target without requiring further design iteration by a researcher.

DeepMind reported that AlphaProteo achieved higher experimental binding success rates than existing design methods across all seven target proteins it was tested against, including the SARS-CoV-2 spike protein and several disease-associated targets such as PD-L1, IL-7Rα and VEGF-A. Across successful designs, the company said binding strength improved by a factor of roughly 3 to 300 compared with the best prior methods, with an average improvement around tenfold. On VEGF-A, a protein implicated in cancer and diabetic eye disease, DeepMind said previous computational design approaches had failed to produce any working binder at all, making it a case where the new system succeeded rather than merely improved on existing results.

The claims came from DeepMind’s own wet-lab validation rather than independent replication, and the blog post did not disclose full details of the model architecture or training procedure, limiting outside researchers’ ability to assess the results without DeepMind’s cooperation — a familiar tension in the “AI for science” announcements labs had increasingly used to demonstrate frontier-model value beyond chatbots and benchmarks. DeepMind said it planned further validation and eventual academic access to support drug discovery, biosensor design and research into pest-resistant crops, positioning the tool as an early step toward automating protein-binder design that historically required years of laboratory iteration.