How to Spend Your Oracle Budget: Practical Guidance for Protein Structure Prediction Models

2026-08-12

AI Summary

This study addresses the limited reliability of foundation models in protein structure prediction for certain targets and the need to efficiently utilize high-cost biological oracles under constrained budgets. It introduces the O3 method into protein structure prediction for the first time and systematically evaluates multiple oracle budget allocation strategies—including FK-steering, DPO, Best K-of-N, and O3—by integrating off-the-shelf optimizers within the generative model’s latent subspace for structural refinement. Experiments on calmodulin (1CLL) and E. coli aspartate transcarbamoylase (9EEH) demonstrate that O3 performs best under low oracle budgets, while FK-steering and DPO excel with higher budgets, with no single approach universally dominating. The work establishes the first comprehensive benchmark and practical recommendation framework tailored to oracle budget constraints in protein structure prediction.

Abstract

Foundation models for protein structure prediction remain unreliable on certain targets. External oracles can flag and correct these failures, but biological oracles are expensive, making oracle budget a critical constraint. Existing guidance methods, such as FK-steering, DPO, and Best K-of-N sampling, differ in how they spend this budget, yet no systematic comparison exists to guide method selection. To bridge this gap, we benchmark these methods alongside the recently proposed Optimisation Over Outputs (O3), which applies off-the-shelf optimisers within a generative model's latent subspace. We extend the usage of O3 to protein structure prediction models. Overall, our work provides the first practical reference for oracle budget-aware guidance. Our evaluation on two protein targets, calmodulin (1CLL) and E. coli aspartate transcarbamoylase (9EEH), reveals that no single method consistently dominates across all budgets and oracles. Specifically, O3 proves most effective at low oracle budgets, while FK-steering and DPO demonstrate improved performance as the budget increases. We distil these findings into actionable recommendations for practitioners operating under real-world oracle-budget constraints.

Problem

• protein structure prediction
• oracle budget
• guidance methods
• foundation models
• biological oracles

Innovation

• oracle budget
• protein structure prediction
• Optimisation Over Outputs
• guidance methods
• foundation models