Generative 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. Since yield varies across these conditions, steering unpromising setups wastes compute. The core insight is that generative condition selection itself becomes a resource-allocation problem at inference time.

Dynamic budget allocation across conditions

The authors propose two dynamic budget allocators that steer and allocate compute across configurations concurrently. The first is a joint Global Sequential Monte Carlo (SMC) formulation that maintains a single system of particles across all candidate conditions (scaffolds or epitopes). Unlike independent steering of separate populations, Global SMC evaluates raw in-loop ipTM and accumulates reward increments across the entire particle population. During global systematic resampling, the computational budget emerges dynamically as unpromising particles are extinguished and successful ones replicate across condition boundaries. The second approach adapts Batched Successive Elimination (BaSE), formulating condition selection as a multi-armed bandit problem. Configurations (arms) are steered independently, with the total computational budget allocated across successive rounds. The algorithm ranks arms by the 90th percentile of their terminal validation scores (AF2-M ipTM) and enforces a deterministic 50% elimination rate at the end of each round, distributing per-round budgets evenly among surviving arms.

Both methods initialize with a uniform prior that distributes the starting computational budget equally across all candidate configurations. The key distinction is that Global SMC targets a cumulative reward path through weight recursion, while BaSE uses a rank-based elimination schedule. Ablation results show that condition allocation alone provides benefits over the vanilla baseline, and that steering and allocation are additive: unsteered SH yielded fewer passing designs than steered SH, demonstrating complementary contributions.

Empirical results: 5× yield improvement per GPU-hour

Empirically, these dynamic allocators increase the yield of designable candidates per GPU-hour nearly 5× without sacrificing diversity across distinct therapeutic targets and generative configurations. Across three targets—PD-L1, IL-7Rα, and BHRF1—adaptive global SMC steering achieved nearly a 5× increase in designable backbones per GPU-hour over the vanilla baseline (rising from 1.41 to 6.82 for BHRF1, and 2.21 to 10.81 for PD-L1). Decoupled SH steering similarly yielded a 2.4–3.1× throughput improvement. The increase in yield did not sacrifice global conformational diversity across independent runs; steering collapses CDR-H3 diversity within individual trajectories but recovers structural variation across independent seeds by exploring distinct high-reward conformational basins.

Oracle divergence and reward hacking

While joint SMC allocation maximizes raw yield, use of orthogonal oracles for steering and evaluation exposes a vulnerability to reward hacking. In PD-L1 binder design, SMC allocated 53% of particles to a framework derived from high ESMFold2 ipTM but yielded zero terminal AF2-M passes due to pose divergence. Decoupled SH mitigates this issue by using orthogonal terminal validation scores for rank cuts. The authors note that both ipTM metrics remain in silico surrogates; high scores do not guarantee in vitro affinity or stability, leaving optimization vulnerable to model-specific artifacts. When serial rounds are feasible, SH is preferable for robustness against in-loop reward hacking, whereas SMC is suited for settings where in-loop rewards are reliable or offline validation is costly.

Practical implications

The framework converts static scaffold and epitope constraints into active optimization variables, enabling sample-efficient de novo nanobody design. Powered by dynamic resource allocation paired with inference-time steering, researchers can distribute compute toward the most promising conditions rather than wasting budget on unpromising upfront commitments. The code and data are released under the CC BY 4.0 license.