Rescuing true protein binders from AI hallucinations via zero-shot, ensemble-driven statistical physics scoring

The advancement of deep generative models has facilitated de novo protein and antibody design, yet translation to experimental success is hindered by a high generation rate of structural decoys. Current affinity predictors and standard structural confidence metrics fail to reliably distinguish these AI hallucinations from true binders. Here, we present Sipobe-PPA, an affinity ranking framework that conceptualizes interacting protein interfaces as pseudo-ligands, evaluating them through an AI-driven statistical physics forcefield. Because this forcefield is trained exclusively on small-molecule interactions, Sipobe-PPA acts as a zero-shot physical evaluator for protein-protein interfaces, preventing the framework against the data leakage and memorization pitfalls that affect models trained directly on protein complex datasets. To capture the structural plasticity of binding interactions, Sipobe-PPA employs a conformational ensemble strategy, computing interaction scores across multiple AlphaFold3(AF3)-predicted structural states. Benchmarking on decoy-rich de novo datasets-including Bindcraft, Boltzgen, and the Germinal antibody dataset-demonstrates the significant improvement offered by this approach. In a real-world pipeline scenario simulating wet-lab constraints (pre-filtered by AF3 ipTM > 0.8 and pLDDT > 80), Sipobe-PPA achieved an 80% Hit Rate within its Top 5 predictions across the combined dataset, compared to 0% for physical baselines like Rosetta-{Delta}G. Notably, our structural ensemble averaging outperformed single-structure scoring, highlighting the necessity of modeling prediction diversity. By maximizing top-tier hit rates across diverse nanobody and de novo targets, Sipobe-PPA provides a scalable screening paradigm that bridges the gap between computational generation and wet-lab viability.