Google DeepMind has published results for AlphaFold 3, the latest iteration of its protein structure prediction system, which now accurately predicts how small drug molecules bind to their protein targets — a capability that could shave years off the drug discovery pipeline.
In blind benchmarks conducted with pharmaceutical partners, AlphaFold 3 predicted binding poses with 89% accuracy (RMSD < 2Å for the predicted binding mode), compared to 45% for the previous best computational method. This is significant because determining how a drug molecule fits into a protein's binding pocket is one of the most expensive and time-consuming steps in early-stage drug development.
Pharmaceutical Partnerships
DeepMind has partnered with Eli Lilly, Novartis, and several biotech startups to validate AlphaFold 3 in real-world drug discovery programs. Early results from Eli Lilly show that the system identified viable drug candidates for a previously “undruggable” cancer target in 3 months — a process that typically takes 18-24 months.
The paper, published in Nature, details the model’s architecture, which extends AlphaFold 2’s protein prediction capabilities with a new “molecular docking” module trained on 12 million experimentally determined protein-ligand structures.
Availability
AlphaFold 3 is available through Google Cloud’s Vertex AI for pharmaceutical researchers. The system processes a typical drug target in under 4 hours on a single TPU v5 pod, compared to weeks of traditional computational docking approaches.