Pitfalls of protein interaction predictors for viral spillover risk prediction

Zoonotic spillover requires viral proteins to achieve molecular compatibility with human host proteins, yet systematic prediction of this remains out of reach. Protein-protein interaction (PPI) models offer a potential path forward, but face critical limitations when applied to viral-host interactions. To directly assess their utility for spillover risk prediction, we introduce a host ortholog ranking benchmark in which models must score a viral protein's binding to human orthologs of known interacting partners above orthologs from non-host species. We find that current PPI models largely fail at this task. To understand the source of this failure, especially for structure-based modeling, we systematically benchmark these models on viral-host complex structure prediction--revealing that viral-host PPIs are substantially harder than general protein complex prediction, due to sparse co-evolutionary signal, shallow viral MSAs, and poor template coverage. We demonstrate how Boltz-2 perturbations - such as dropout, noise, or structure-based recycling, raise structure prediction performance on a subset of viral receptor complexes, and interrogate the importance of model components like distogram contacts or MSA conservation signal. We also fine-tune Boltz-2 specifically for unpaired MSA complex prediction. Finally, we ask whether high-quality structure predictions can rescue ortholog ranking: models with high structure prediction accuracy of single orthologs can distinguish host from non-host orthologs across the benchmark, but others cannot. Moreover, the false positive rate of model prediction also leads to over-prediction of likely viral receptors. These results reveal fundamental gaps in current PPI models for viral-host compatibility prediction and motivate purpose-built approaches for spillover risk assessment.