MIT and DeepMind Joint Study Achieves 99.1% Accuracy in Predicting Protein-Molecule Binding

A collaborative team from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and Google DeepMind published a landmark paper this week introducing ProBind 3, a model that predicts drug-protein binding affinity with 99.1% accuracy on the PDB benchmark — shattering the previous record of 91.4% held by the DeepMind AlphaFold 3 system.

The implications for pharmaceutical research are significant. Drug-protein binding prediction is one of the most time-consuming steps in early-stage drug discovery; getting it wrong means months of wasted lab work. ProBind 3's leap in accuracy means researchers can virtually screen thousands of candidate molecules before stepping into a wet lab, potentially compressing discovery timelines from years down to months.

The model was trained on a dataset of 12 million protein-ligand complexes, leveraging a new graph neural network architecture that treats both the protein pocket and the drug molecule as interacting 3D point clouds. DeepMind contributed its TPU v6 compute infrastructure; MIT provided the dataset curation and experimental validation using high-throughput crystallography.

The team has released ProBind 3 weights under an open research license. Pharmaceutical partners including Novartis and Roche have already signed agreements to integrate ProBind 3 into their discovery pipelines.