Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design

The gist The authors propose two dynamic budget allocators, a joint Global Sequential Monte Carlo (SMC) formulation and a decoupled approach adapting Batched Successive Elimination (BaSE), to increase the yield of designable candidates per GPU-hour in de novo nanobody design by treating generative condition selection as an inferencetime resource-allocation problem.

How it works

The paper addresses the bottleneck in de novo nanobody design where upfront commitment to conditioning choices like scaffold identity or target epitope wastes compute because yield varies across these conditions. To solve this, they recast condition selection as an inferencetime resource-allocation problem.

The proposed methods include two dynamic allocators:

1. Joint Global Sequential Monte Carlo (SMC): This formulation maintains a single, joint system of N particles and targets the reward-tilted joint distribution π(k, x0:T) ∝ ωk qk(x0:T) exp(Φ(x0:T)/τ), where Φ(x0:T) is the additive reward path functional. This approach targets a cumulative reward path, and emergent budget allocation occurs during global systematic resampling as unpromising particles are extinguished and successful ones replicate across condition boundaries.

2. Decoupled Allocation via Successive Halving (SH): This method formulates condition selection as a multi-armed bandit problem using the Batched Successive Elimination (BaSE) framework. It modifies standard BaSE by distributing per-round budgets evenly among surviving arms, ranking arms by the 90th percentile of their terminal validation scores, and enforcing a deterministic 50% elimination rate at the end of each round.

Steering Mechanisms

Inference-time steering methods like Feynman-Kac (FK) steering are applied to sample from a reward-tilted distribution by actively cloning promising trajectories and culling low-scoring ones. In the context of fixed-scaffold nanobody generation, this involves interleaving standard denoising steps with guidance events that evaluate particle quality, discarding unpromising states and replicating successful ones.

Steering parameters were empirically optimized for each experiment, utilizing a population of N = 16 particles for standard steered runs and K × 16 particles in joint allocation experiments where K is the total number of candidate conditions. The steering temperature, τ ∈ [0.1, 0.25], was determined via a grid search.

Results and Discussion

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. For example, adaptive global SMC steering achieved nearly a 5× increase in designable backbones per GPU-hour over the vanilla baseline for BHRF1, rising from 1.41 to 6.82. Similarly, decoupled SH steering yielded a 2.4–3.1× throughput improvement.

The paper demonstrates that while steering collapses paired CDR-H3 Cα RMSD within individual trajectories, this loss of variation is localized to single runs. However, across independent seeds, steering recovers structural variation by exploring distinct, high-reward conformational basins. This suggests that steering does not inherently sacrifice global diversity; rather, it highlights the need to distribute compute across multiple starting conditions to sample the viable paratope space.

Furthermore, while joint SMC allocation maximizes raw yield in the PD-L1 binder design experiment, it exposes a vulnerability to reward hacking because SMC allocated 53% of its particles to the 8z8v-derived framework due to 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, making SH preferable when serial rounds are feasible.

Validation and Conclusion

The post-hoc designability validation pipeline evaluates candidates using three metrics: inter-chain predicted alignment error (ipAE) 90, and the self-consistent root-mean-square deviation of the binder (binder scRMSD) < 1.5 Å. The study concludes that dynamic resource allocation paired with inference-time steering converts static scaffold and epitope constraints into active optimization variables, increasing in silico nanobody yield without sacrificing structural diversity. Adaptation of these allocators to wet-lab-correlated oracles represents a primary direction for future work. The authors also note that high ipTM scores do not guarantee in vitro affinity (KD) or stability, leaving optimization vulnerable to model-specific artifacts.

Acknowledgements

The authors thank Mike Schmidt and Piotr Bobrowicz for their scientific support and critical review of the manuscript. They also thank Esteban Gea for his assistance with the computational infrastructure used in early experiments.

References

[1] Hannes Stark, Felix Faltings, MinGyu Choi, Yuxin Xie, Eunsu Hur, Timothy O’Donnell, Anton Bushuiev, Talip Uçar, Saro Passaro, Weian Mao, Mateo Reveiz, Roman Bushuiev, Tomáš Pluskal, Josef Sivic, Karsten Kreis (2025). BoltzGen: Toward universal binder design. bioRxiv.

[2] Jarrid Rector-Brooks et al. (2026). General multimodal protein design enables dna-encoding of chemistry.

[3] Kieran Didi et al. (2026). Scaling atomistic protein binder design with generative pretraining and test-time compute. In The Fourteenth International Conference on Learning Representations, 2026.

[4] Raghav Singhal et al. (2025). A general framework for inference-time scaling and steering of diffusion models. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 55810–55827. PMLR, 2025.

[5] Marta Skreta et al. (2025). Feynmankac correctors in diffusion: Annealing, guidance, and product of experts. arXiv preprint arXiv:2503.02819, 2025.

[6] Bradley Efron (2011). Tweedie’s formula and selection bias. Journal of the American Statistical Association, 106(496):1602–1614, 2011.

[7] Erik Hartman et al. (2025). Controllable protein design with particle-based Feynman–Kac steering. arXiv preprint arXiv:2511.09216, 2025.

[8] Kiwoong Yoo et al. (GEM 2026). Search-based inference-time scaling for all-atom protein binder design. In ICLR 2026 Workshop on Generative and Experimental Perspectives for Biomolecular Design (GEM), 2026.

[9] Nathaniel R Bennett et al. (Nature, 2026). Atomically accurate de novo design of antibodies with RFdiffusion. Nature, 649(8095):183–193, 2026.

[10] Luis S. Mille-Fragoso et al. (Nature Biotechnology, 2026). Efficient generation of epitope-targeted antibodies with Germinal. Nature Biotechnology, 2026.

[11] Cécile Vincke et al. (Journal of Biological Chemistry, 2009). General strategy to humanize a camelid single-domain antibody and identification of a universal humanized nanobody scaffold. Journal of Biological Chemistry, 284(5):3273–3284, 2009.

[12] Martin A. Rossotti et al. (The FEBS Journal, 2022). Immunogenicity and humanization of single-domain antibodies.

1. textbf Dynamic Budget Allocation for Condition Selection in Generative Design (Global SMC): The system can address upfront commitment to conditioning choices like scaffold identity or target epitope by using a joint Global Sequential Monte Carlo (SMC) formulation that steers and allocates compute across configurations concurrently. This allows the AI to sample from a distribution that targets a cumulative reward path, where the budget is dynamically allocated as unpromising particles are extinguished and successful ones replicate across condition boundaries, leading to an increase in yield of designable candidates per GPU-hour.

2. textbf Decoupled Condition Selection via Multi-Armed Bandits (SH): The system can manage the condition selection bottleneck by employing a decoupled approach adapting Batched Successive Elimination (BaSE) where configurations are evaluated by steering independently and ranking them using the 90th percentile of their terminal validation scores (AF2-M ipTM). This provides a robust method for selecting promising scaffolds or epitopes, as demonstrated by the finding that decoupled SH steering yields 2.4–3.1× throughput improvement.

3. textbf Adaptive Steering Temperature Control: The system can optimize the exploration-exploitation trade-off by using an adaptive policy dynamically triggered by weight degeneracy (ESSm ≤ 0.5N) for resampling triggers, as opposed to a fixed cadence. This allows the sampler to be more efficient when trajectories are highly promising, maximizing throughput while maintaining diversity, as shown by the results showing that adaptive global SMC steering achieved nearly a 5× increase in designable backbones per GPU-hour.

4. textbf Robustness Against Reward Hacking: The system can mitigate the risk of reward hacking inherent in in-loop metrics by employing decoupled allocation strategies. By ranking arms using orthogonal terminal validation scores for rank cuts, the AI ensures that optimization is guided by external, more reliable metrics, as noted by the conclusion that SH is preferable when serial rounds are feasible.

5. textbf Explicit Diversity Management: The system can ensure global structural exploration despite steering constraints by leveraging multiple independent runs. The paper demonstrates that steering does not inherently sacrifice global diversity; rather, it highlights the need to distribute compute across multiple starting conditions to sample the viable paratope space, as evidenced by the recovery of distinct, high-reward conformational basins across independent seeds.

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. We recast generative condition selection as an inference-time resource-allocation problem. We propose two dynamic budget allocators: a joint Global Sequential Monte Carlo (SMC) formulation that steers and allocates compute across configurations concurrently, and a decoupled approach adapting Batched Successive Elimination (BaSE). Empirically, these dynamic allocators increase the yield of designable candidates per GPU-hour nearly 5 times without sacrificing diversity across distinct therapeutic targets and generative configurations.