Design-CP Boosts Protein Nanoparticle Design with Context Parallelism

Generative protein models, particularly all-atom ones, theoretically allow for the design of large multimeric protein complexes by modeling all chains simultaneously. However, their quadratic token- and atom-pair representations quickly exhaust single-GPU memory as the number of chains and residues increases. To overcome this limitation, researchers have developed Design-CP. Design-CP introduces two context-parallel (CP) inference strategies for RFdiffusion 3: 1D row-sharding and 2D grid sharding with ring attention. These strategies distribute the computationally intensive quadratic activations across a multi-GPU mesh while preserving pretrained model weights. The study characterizes their scaling performance when sampling icosahedral assemblies, showing that the maximum feasible asymmetric subunit size grows with the expected square-root trend in GPU count, with 2D sharding achieving better wall-clock scaling. The research also demonstrates how strong point-group symmetry constraints make CP immediately usable for end-to-end, all-atom design of icosahedral nanoparticles, yielding favorable structural and interface metrics. Finally, it showcases octahedral nanoparticle design on a small cluster of workstation-grade 16GB GPUs, illustrating Design-CP's potential to democratize large-assembly protein design.