Isomorphic Labs, the Alphabet-backed AI drug discovery company led by Nobel laureate Sir Demis Hassabis, has unveiled its Drug Design Engine (IsoDDE) and the results are striking enough that the story rocketed to the top of Hacker News on July 18. The engine more than doubles the accuracy of AlphaFold 3 on the most difficult protein-ligand structure prediction benchmarks, predicts binding affinities that surpass gold-standard physics-based methods, and can identify novel drug binding pockets using nothing more than a protein's amino acid sequence as input. For founders building in healthtech and biotech, this is the moment the promise of AI-driven drug design moves from research headlines into a production-grade computational workflow.
What the Drug Design Engine Does That AlphaFold Could Not
AlphaFold 3, released by Google DeepMind and Isomorphic Labs in 2024, transformed protein structure prediction. Over 3 million researchers across 190 countries have used it. But structure prediction alone is not drug design. Knowing the 3D shape of a protein tells you where a pocket might be, but it does not tell you which molecules will bind strongly to it, how a protein will reshape itself when a molecule docks, or whether a hidden pocket exists that only opens under specific conditions. IsoDDE addresses all three gaps in a unified system. On the Runs N' Poses benchmark, a test specifically designed to measure generalization to unfamiliar protein-ligand systems, IsoDDE more than doubles AlphaFold 3's accuracy on the hardest cases. It models complex events like induced fits, where a protein reshapes itself to accommodate a bound molecule, and cryptic pockets, those completely invisible in the absence of a binding ligand. These are precisely the phenomena that determine whether a promising molecule becomes a real drug.
Binding Affinity Predictions That Beat Physics-Based Methods
One of the most important capabilities IsoDDE introduces is binding affinity prediction. Drug development hinges on knowing how strongly a candidate molecule will stick to its target. Traditional approaches fall into two camps: machine learning models that are fast but inaccurate on unfamiliar chemistry, and physics-based methods like free energy perturbation (FEP) that are accurate but prohibitively slow and expensive, requiring experimental crystal structures as a starting point. IsoDDE surpasses all existing deep-learning methods by a wide margin on three public benchmarks: the FEP+ dataset, the OpenFE benchmark, and the CASP16 blind binding affinity prediction task. Remarkably, it can exceed the performance of physics-based FEP methods despite not requiring experimental crystal structures at all. This means researchers can rank and optimize candidate molecules in silico at speeds that were previously impossible, collapsing what used to be months of experimental screening into hours of computation.
The Cereblon Breakthrough: Finding Hidden Pockets from Sequence Alone
The most striking demonstration of IsoDDE's capabilities involves the protein cereblon. For 15 years, researchers believed there was one principal way to drug cereblon: through the thalidomide-binding pocket. A 2026 study by Dippon and colleagues experimentally discovered a second, previously hidden pocket that works allosterically, meaning it modulates the protein from a distance. IsoDDE recapitulated this discovery using only cereblon's amino acid sequence as input. It predicted both the known pocket and the novel cryptic site without any information about what molecules might bind there. When the actual ligands were specified, it correctly docked them into their respective pockets in the right orientation. This capability for blind pocket identification approaches the accuracy of experimental techniques like fragment soaking, which require significant time, cost, and wet-lab work. IsoDDE does it in seconds on a computer. For any startup working on a first-in-class drug target with no structural annotation, this capability alone could reshape the early discovery pipeline.
What This Means for Founders Building in AI-Driven Biotech
IsoDDE represents a step change in what AI can do for drug discovery, and it carries direct implications for founders. First, the bar for AI capability in biotech just moved. Any startup claiming AI-driven drug design will now be measured against IsoDDE's benchmark results. Second, the compute-to-discovery ratio is shifting dramatically. If binding affinity predictions that once required weeks of physics-based computation can now run in seconds, the bottleneck in early-stage drug development moves from computation to biological question-asking. Third, Isomorphic Labs' approach of building a unified system rather than a collection of point solutions suggests that vertical AI platforms will increasingly outcompete single-model tools. Founders building in healthtech should watch which other domains could benefit from this kind of integrated AI design engine. Enzyme engineering, synthetic biology, and protein therapeutics are all candidates. The era of treating AI as a nice-to-have add-on in drug discovery is over. IsoDDE signals that AI-native design is becoming the default workflow, not the experimental one.