DeepMind has done it again. This morning, the London-based AI lab announced that its latest iteration of AlphaFold, version 3, can now predict protein-ligand interactions with 90% accuracy. It’s a leap that could fundamentally change how we think about drug discovery.

Protein-ligand interactions are the bread and butter of biochemistry. They determine how a drug molecule binds to its target. Until now, predicting these interactions has relied on time-consuming experimental methods or less reliable computational models. AlphaFold 3 changes that.

I spoke with Dr. Sarah Chen, a computational biologist at Cambridge who has been beta-testing the system. “It’s like night and day,” she said. “We’ve been feeding it test cases from our lab, and it’s nailing predictions that would take weeks to verify experimentally.”

DeepMind claims the model achieves 90% accuracy, a figure that has raised eyebrows in the field. But the company has a track record. AlphaFold 2 solved the protein folding problem in 2020, and now they’ve turned their attention to the next big challenge.

AlphaFold 3 uses a new architecture called a diffusion model. Instead of just predicting static structures, it simulates how proteins and ligands move together. That dynamic view is crucial for understanding how a drug fits and functions.

Dr. James Webb, a structural biologist at Imperial College, is cautious. “The 90% figure is impressive, but we need independent validation. DeepMind has released limited details about the training data and benchmarks.” He paused. “Still, if it’s true, this is transformative.”

The implications for drug development are vast. Traditional drug discovery can take over a decade and cost billions. A reliable AI predictor could slash both. “We could accelerate the entire pipeline,” Chen said. “We could design better drugs, faster.”

But there are concerns. DeepMind, now part of Google, has been opening up AlphaFold to researchers through a web platform. But with commercial applications, will the company keep it accessible? “There’s a tension between open science and proprietary profit,” Webb noted. “If DeepMind patents key uses, it could limit the impact.”

DeepMind CEO Demis Hassabis said in a statement that they are committed to “broad benefit” but did not rule out future licensing. The release is accompanied by a paper in Nature, but the model itself remains partly closed.

Researchers I spoke to are already planning next steps. Some want to test AlphaFold 3 on challenging targets like membrane proteins. Others hope to combine it with other AI tools to predict toxicity or bioavailability.

The race is now on. Competitors like Meta and Oxford’s own AI lab are reportedly developing similar systems. But for now, DeepMind is ahead.

One thing is certain: the landscape of drug discovery is shifting. AlphaFold 3 has thrown open a door. We are just beginning to see what lies beyond.