Google DeepMind previews next-gen AlphaFold that folds DNA and drugs

Predicting how a single protein folds was a massive scientific leap, but proteins inside our bodies never act alone. To actually design a cure, you need to see how drugs, DNA, and other biochemical machinery latch onto them in real time.

Google DeepMind and its drug discovery spin-off Isomorphic Labs just unveiled a major progress update on the next generation of AlphaFold. Moving far beyond isolated protein chains, the new model expands its coverage to ligands, DNA and RNA nucleic acids, and structures with post-translational modifications.

According to DeepMind, the latest system can generate structure predictions for nearly all molecule types found in the Protein Data Bank, frequently achieving atomic accuracy.

Ditching the rigid constraints of traditional docking

For decades, the standard way to model drug-protein interactions has been computational docking. The catch with docking is that it requires a known, rigid reference structure and a researcher's best guess as to where the binding pocket is.

DeepMind's new model sidesteps those crutches entirely. It outperforms top docking methods without needing an existing reference structure or any prior hint about where the pocket sits, opening the door to predict interactions for completely uncharacterized proteins.

Even better, the model predicts the positions of all atoms jointly. This allows it to capture the subtle, flexible morphing that happens when proteins and nucleic acids hug other molecules, something rigid docking tools could never reproduce.

DeepMind showcased several tough biomedical targets where the model matched lab-bench experiments almost down to the atom. These include the notorious cancer target KRAS involving molecular glue, the cancer drug candidate PORCN, and the CRISPR-associated gene editing scissors CasLambda bound to DNA and guide RNA.

Why it matters

Over 1.4 million researchers in 190 countries have relied on AlphaFold, but designing medicine has still required a patchwork of legacy tools to simulate actual drug binding. If this next-generation model delivers as promised, drug hunters and synthetic biologists will finally be able to simulate multi-molecule drug targeting directly inside a single neural network.

For the biotech sector, this shifts digital biology from passive structural mapping to active molecular design. DeepMind is clearly positioning Isomorphic Labs to lead commercial drug pipelines, turning an academic triumph into a proprietary drug engine that puts legacy computational chemistry platforms on defense.

We think expanding into full biomolecular complexes is the single smartest move DeepMind could make right now. But remember that predicting binding affinity in a simulation is only half the battle. Molecules still have to survive human metabolism, bypass toxicity barriers, and show actual efficacy in clinical trials, so don't expect wet-lab validation to disappear anytime soon.