AlphaFold 4 Predicts Drug-Protein Binding with 94% Accuracy

Google DeepMind has released AlphaFold 4, the newest version of its protein structure prediction system, and it can now model how drug molecules bind to their protein targets — a critical step in designing new medicines.

Published in Nature, the model was tested on 500 known drug-target pairs and correctly predicted binding affinity in 94% of cases. That's a meaningful jump from AlphaFold 3, which could map protein shapes but struggled with the dynamic, flexible interactions between proteins and small molecules that determine whether a drug actually works.

AlphaFold 4 uses a diffusion-based architecture, which lets it simulate how drug candidates physically conform to their targets — something like watching a key fit into a lock, but at atomic scale.

"We can now go from a protein target to a ranked list of drug candidates in hours instead of months," said Demis Hassabis, DeepMind's CEO.

Pharmaceutical companies are already testing the system. Eli Lilly ran a six-month pilot and reports a 40% reduction in time spent screening lead candidates for its oncology pipeline. Novartis is using AlphaFold 4 for early-stage research on neurodegenerative diseases.

But some researchers are tempering optimism. Dr. Maria Rodriguez, a computational biologist at Stanford not involved in the work, noted that "predicting binding affinity in silico is not the same as predicting efficacy in vivo. There's a long road between a computational prediction and a drug that works in patients."

The system will be available to academic researchers through Google Cloud, with pharmaceutical partners paying for API access. DeepMind says it plans to release model weights for non-commercial use later this year.

---

This article synthesizes reporting from MIT Technology Review, with verification against the Nature paper methodology and framing of company-reported results as claims rather than independently confirmed outcomes.