BioEvidence lets frozen LLMs match strong specialized protein predictors in zero-shot variant ranking, with performance rising as evidence gets richer
The authors introduce BioEvidence, a training-free and model-agnostic interface that converts structural and evolutionary information from standard biological tools into compact evidence for frozen large language models, and on the ProteinGym benchmark they observe evidence scaling: performance improves as evidence becomes richer, structural and evolutionary evidence each help and combining them yields further gains, mismatching the same evidence to the wrong variants degrades performance below the no-evidence baseline, and the interface enables zero-shot ranking to reach strong specialized protein predictors on matched evaluations while the improvement persists on data released after the model's knowledge cutoff.