New Design-CP method scales protein nanoparticle design across GPUs

Designing large multimeric protein complexes like icosahedral nanoparticles using all-atom generative models has been bottlenecked by GPU memory—quadratic token- and atom-pair representations quickly exceed single-GPU limits. A new paper from Lorenzo Tarricone, Helen E. Eisenach, Aiko Muraishi, and Charlotte M. Deane (University of Oxford) introduces Design-CP, a set of context-parallel inference strategies for RFdiffusion 3 that distribute these massive activations across multiple GPUs while preserving pretrained weights. The two strategies—1D row-sharding and 2D grid sharding with ring attention—allow researchers to scale up the number of chains and residues modeled without hitting memory walls.

The team characterized scaling on icosahedral assemblies, showing that the maximum feasible asymmetric subunit (ASU) size grows with the expected square-root trend as GPUs increase, and that 2D sharding achieves significantly better wall-clock scaling than 1D. Strong point-group symmetry constraints make CP usable out of the box for end-to-end, all-atom design, yielding favorable in silico structural and interface metrics. Crucially, they demonstrated octahedral nanoparticle design on a small cluster of workstation-grade 16GB GPUs, proving that large-assembly protein design is now practical on accessible hardware. Accepted at the ICML 2026 Workshop on Generative and Agentic AI for Biology, Design-CP offers a concrete path toward democratizing computational protein design.