Anthropic has released material demonstrating how the Claude Mythos 5.2 model solved the problem of predicting the 3D structure of multi-chain AAV capsids. Instead of attempting to fold 60 protein chains at once, the AI wrote code that decomposed the problem into smaller groups, achieving single-atom accuracy in 7 to 58 minutes on a single GPU.
The problem of 60 protein chains and the limits of traditional modeling
Designing safe and effective viral vectors for gene therapy has long been hampered by computational constraints. A key component of these therapies are the capsids of adeno-associated viruses, commonly known as AAV (Adeno-associated virus). A capsid is essentially a protein shell designed to safely transport genetic material directly into a patient's cells. The difficulty in predicting its 3D structure lies in the scale of the problem: the AAV shell consists of 60 identical copies of a single protein chain, arranged into a precise, symmetrical icosahedron.
When we attempt to feed such a complex structure into standard, open-source protein folding prediction models, we immediately hit a wall. Trying to calculate the interactions between all 60 chains simultaneously exceeds the memory and computational capacity of most individual GPU accelerators. Instead of a fluid result, we receive memory overflow errors or complete distortion of the geometry of the connections between individual monomers.
In a video published on November 9, 2026, on the official Claude channel, an approach was presented that bypasses this problem through software engineering and symmetry algebra. As stated in the material and on the accompanying Anthropic documentation page, the Claude Mythos 5.2 system did not attempt to solve the entire system in one massive calculation. Instead, it prepared a coordination script that divided the task into smaller, overlapping subgroups of chains.
Overlapping pentamer strategy: how Claude Mythos 5.2 bypassed GPU barriers
Instead of treating 60 chains as one giant object, the code generated by Claude Mythos 5.2 isolated smaller assemblies – adjacent groups of five chains (so-called pentamers). Each pentameric fragment partially overlapped with the next, which allowed for the maintenance of interaction continuity at the protein interface boundaries.
It is worth clarifying one important point to avoid misconceptions: the AI model itself does not "discover" the symmetry geometry from scratch. Information about the capsid symmetry was provided to the system from the outside as a strict geometric constraint. The task of the algorithm written by Claude Mythos 5.2 was to use this external symmetry to properly position individual pentamers in 3D space and perform iterative optimization of intermolecular interactions.
The test utilized two established, open-source folding models: OpenFold3-preview2 and ESMFold2. Claude Mythos 5.2 acted as the architect of the entire process—it wrote the complete control code, defined optimization loops, set physical parameters, and managed the simulation run on a single graphics card.
Simulation workflow in OpenFold3-preview2: from noise to 1.2 Å
The presented recording shows a detailed workflow of the AAV2 viral capsid folding simulation using the OpenFold3-preview2 model. This process involves the gradual refinement of the structure and noise reduction—for example, at the 135th step, the noise level was 2.8, and by the final 200th step, it dropped to 0.1.
The entire simulation in OpenFold3-preview2 was scheduled for 200 computational steps and took 58 minutes on a single GPU. During the simulation, one could observe the gradual ordering of the structure:
- Chain and interface optimization: Each simulation step gradually refines the arrangement of all 60 chains and the alignment of dimer interfaces.
- Disabling the SITE TETHER mechanism: At the 120th step of the simulation, the auxiliary position-binding mechanism (SITE TETHER) is disabled.
- Structure stabilization: In subsequent steps, further noise reduction occurs until the mature capsid form is achieved at the 200th step.
The obtained result was compared with experimental data derived from the X-ray structure. The Cα RMSD (root-mean-square deviation of alpha carbon positions) for the predicted AAV2 capsid was only 1.2 Å (angstroms). From the perspective of physical chemistry, this error is within the range of a single atom's diameter.
Comments