Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design
The paper’s claim is a nearly 5x gain in designable nanobody candidates per GPU-hour by shifting compute away from weak scaffold and epitope choices during generation.
The authors frame de novo nanobody design as a resource-allocation problem at inference time, not just a sequence-generation problem. They propose two allocators: a joint Global SMC method and a decoupled BaSE-style approach. Both steer generation across candidate configurations instead of committing upfront to one scaffold or target epitope. The reported efficiency gain comes without sacrificing diversity across therapeutic targets and generative setups. ArXiv · AI/CL/LG's note
The authors frame de novo nanobody design as a resource-allocation problem at inference time, not just a sequence-generation problem. They propose two allocators: a joint Global SMC method and a decoupled BaSE-style approach. Both steer generation across candidate configurations instead of committing upfront to one scaffold or target epitope. The reported efficiency gain comes without sacrificing diversity across therapeutic targets and generative setups. ArXiv · AI/CL/LG's note
score 5