Self-Improvement Imitation with Biologically Guided Search for Protein Design Under Oracle Budgets

Researchers introduce SILO, a self-improvement imitation framework for protein design that optimizes protein sequences under limited evaluation budgets. The method combines hierarchical editing, stochastic beam search, and active learning to outperform existing reinforcement learning and generative approaches across multiple protein fitness landscapes.

SILO addresses a fundamental challenge in computational biology: designing proteins with desired functions when oracle evaluations are expensive and limited. Protein sequence optimization involves searching through astronomically large combinatorial spaces—with roughly 20 amino acid choices at each position across hundreds of positions—making exhaustive evaluation infeasible. Traditional approaches using reinforcement learning or generative models struggle when surrogate models are noisy or when mutations randomly target positions without considering functional importance.

The breakthrough lies in SILO's hierarchical approach and active learning strategy. By decomposing mutations into position selection followed by residue selection, the framework can intelligently prioritize functionally critical regions identified through alanine-scan fitness scoring. The incremental stochastic beam search without replacement generates diverse candidate trajectories efficiently, while a UCB-based ensemble proxy filters for candidates most likely to yield improvements. This combination ensures that expensive oracle evaluations focus on genuinely promising mutations rather than random searches.

The results demonstrate substantial practical impact: SILO achieved best or tied performance across all eight tested protein landscapes, with particularly strong advantages in low-data and noisy conditions where competing methods degrade. This matters for biotechnology development, where protein engineering applications in therapeutics, industrial enzymes, and synthetic biology depend on efficient optimization under real-world constraints.

The framework's robustness under adverse conditions suggests it could accelerate drug discovery pipelines and enzyme engineering projects. Ablation studies confirm that the core innovations—beam search combined with alanine-scan filtering and iterative imitation—drive the gains, providing clarity on which components are essential. Open-source code availability enables rapid adoption across research and industry applications.

• →SILO's hierarchical edit policy and active learning strategy outperform existing methods on all tested protein fitness landscapes
• →The framework maintains competitive performance in low-data and noisy conditions where baseline methods fail
• →Alanine-scan fitness scoring combined with stochastic beam search accounts for majority of performance gains
• →The method avoids expensive value-function estimation by using cross-entropy imitation on oracle-labeled trajectories
• →Open-source implementation enables practical deployment in biotechnology and drug discovery applications