Isomorphic Labs says its Drug Design Engine beats AlphaFold 3 on structure prediction

Isomorphic Labs says its Drug Design Engine more than doubles the accuracy of AlphaFold 3 on a demanding protein-ligand structure-prediction benchmark. The Alphabet-owned drug-discovery company published the claim in a February technical account that has drawn renewed attention this week.

The system, called IsoDDE, is a unified computational drug-design engine that Isomorphic Labs describes as a step beyond AlphaFold 3, the protein-structure model from sibling lab Google DeepMind. IsoDDE predicts small-molecule binding affinities with accuracy the company says exceeds gold-standard physics-based methods on the FEP+, OpenFE and CASP16 benchmarks.

Isomorphic Labs also says IsoDDE can identify novel binding pockets on target proteins from amino-acid sequence alone, and that on antibody-antigen structure prediction it outperforms AlphaFold 3 by 2.3 times and the Boltz-2 model by 19.8 times on high-fidelity predictions, measured by a DockQ score above 0.8.

Every one of those numbers is company-supplied and not independently verified. Isomorphic Labs published the results in a blog-style article rather than a peer-reviewed paper, and did not release the engine for outside testing. Benchmark accuracy in binding-affinity prediction has repeatedly failed to translate cleanly into real drug candidates.

The claims matter because Isomorphic Labs, founded to turn DeepMind’s structure-prediction work into medicines, is racing rival computational-chemistry groups to show that better prediction shortens drug discovery. The proof will be clinical candidates, not benchmark tables.