Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design
Inference-time budget allocators raise designable nanobody yield nearly fivefold per GPU-hour.
The paper recasts scaffold and epitope selection for de novo nanobody design as an inference-time compute-allocation problem rather than a fixed upfront choice. It introduces a joint Global Sequential Monte Carlo allocator and a decoupled Batched Successive Elimination method that steer Feynman-Kac sampling toward high-reward candidates. Empirically, the allocators raise designable-candidate yield per GPU-hour by nearly 5× without reducing diversity across therapeutic targets and generative configurations.
- Scaffold and epitope choice treated as compute allocation
- Global SMC and Batched Successive Elimination allocators proposed
- Nearly 5× more designable candidates per GPU-hour
- Diversity held across therapeutic targets and configurations
From the original article
1 min read at arxiv.orgGenerative molecular design typically relies on over-generation and post-hoc filtering. Inference-time steering methods like Feynman-Kac (FK) steering improve efficiency by biasing trajectories toward high-reward candidates, yet they optimize within a single fixed configuration. In de novo nanobody design, upfront commitment to conditioning choices like scaffold identity or target epitope creates a bottleneck.
By Mason Minot · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.