News · 2026-07-17

Isomorphic Labs Unveils a Drug-Design AI That Beats Physics at Predicting How Drugs Bind

Isomorphic Labs, the drug-discovery company spun out of Google DeepMind, unveiled its Drug Design Engine (IsoDDE) on July 17, 2026, and its central claim is remarkable: on predicting how tightly a candidate molecule binds to a protein, the learned model surpasses not just other AI systems but the physics-based methods that have been the gold standard for decades. In one flagship demonstration, IsoDDE recomputed a drug-binding pocket that took scientists 15 years to discover in the lab -- using only the protein's amino-acid sequence, in seconds.

Key facts

• IsoDDE more than doubles the accuracy of AlphaFold 3 on the hardest, most novel systems (those least similar to training data) on the 'Runs N' Poses' benchmark.
• It surpasses physics-based Free Energy Perturbation on binding-affinity prediction across three public benchmarks, despite not requiring the experimental crystal structures that physics methods need.
• It is 'in production use internally' across Isomorphic's active drug programs, per the company announcement.

To see why this is a step change, start with what came before. AlphaFold, DeepMind's Nobel-recognized system, solved protein structure prediction -- given a sequence, predict the 3D shape. But Isomorphic makes a pointed argument: 'understanding biomolecular structures alone was not sufficient for unlocking real-world drug discovery.' Knowing a protein's shape does not tell you whether a drug will stick to it, how strongly, or where the druggable spots are. IsoDDE is built to close that gap across four capabilities: predicting structures of genuinely novel systems, handling hard biologics like antibody-antigen pairs, predicting binding strength, and finding new pockets.

The binding-affinity result is the most disruptive. For decades, the reliable way to compute how tightly a molecule binds has been Free Energy Perturbation (FEP), a physics simulation that is accurate but slow and needs an experimental crystal structure to start from. Isomorphic reports that IsoDDE beats FEP on three public benchmarks -- a learned model outperforming physics at physics' own game, without needing the experimental input, at a fraction of the time and cost. Think of it as the difference between simulating every water molecule around a drug (physics) and a model that has seen enough examples to predict the outcome directly.

The cereblon story is the one that makes it concrete. Cereblon is a protein that tags damaged proteins for disposal, and for 15 years scientists believed there was essentially one way to drug it -- through the classic thalidomide-binding pocket. Then, in 2026, researchers experimentally discovered a second pocket that was both hidden (invisible without a molecule bound) and located away from the known site. IsoDDE, given only cereblon's sequence and told nothing about the ligands, predicted the location of both the known and the newly discovered hidden pocket -- and then, once molecules were specified, folded them into the correct pockets in the correct orientation. A 15-year experimental result, recomputed on a computer in seconds.

Why it matters: if these results hold up under independent scrutiny, IsoDDE compresses steps of the drug-discovery pipeline that currently take years of wet-lab work into computational predictions -- and, crucially, it generalizes to systems unlike anything in its training data, which is exactly where AlphaFold 3 is weakest. Isomorphic says its drug-design teams already use IsoDDE daily 'to understand unseen structures, identify uncharacterised pockets, and create novel chemical matter.'

The honest caveat is important. Every number here is Isomorphic's own, from its own benchmarks; there is no independent third-party validation of IsoDDE yet. The 'beats physics-based FEP' claim in particular is extraordinary and would need outside replication before the field treats it as settled -- learned models can look great on curated benchmarks and stumble on the messy reality of a live drug program. But the direction is unmistakable, and it fits the broader pattern of AI moving from predicting what molecules look like to predicting what they do.

Primary source, verified: read the paper →

Ground Truth. (2026, July 17). Isomorphic Labs Unveils a Drug-Design AI That Beats Physics at Predicting How Drugs Bind. Ground Truth. https://groundtruth.day/news/isomorphic-labs-drug-design-engine-beyond-alphafold.html

@misc{groundtruth:isomorphic-labs-drug-design-engine-beyond-alphafold,
  title  = {Isomorphic Labs Unveils a Drug-Design AI That Beats Physics at Predicting How Drugs Bind},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/isomorphic-labs-drug-design-engine-beyond-alphafold.html}
}

Topics: ai-for-science · drug-discovery · isomorphic-labs · alphafold · biology