Inference-time scaling of de novo designed proteins is more effective for harder targets

Summary

Inference-time scaling of de novo designed proteins is more effective for harder targets (1). Scaling was executed using beam search, Feynman-Kac steering, and Monte Carlo Tree Search. Surprisingly, all three gave comparable results to one another.

Figures

Ref (1)

See also

• Search algorithms are more important than language model choice for model-guided protein engineering and design

1.

Didi K, Zhang Z, Zhou G, Reidenbach D, Cao Z, Cha S, et al. Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute. 2026; Available from: https://arxiv.org/abs/2603.27950