From Measurement to Confidence: Interpreting Structure Quality Metrics for Predicted Structures Without Experimental Ground Truth

The widespread use of protein structure prediction (PSP) models has increased the need for reliable evaluation methods in the absence of experimental ground-truth data. However, many commonly used structural quality metrics have been developed and calibrated using experimentally determined structures. Can these metrics be directly transferred to predicted structures is a question that this research explores through measurement-inference studies. Here, we evaluate 151 predicted homodimers from AlphaFold-Multimer (AFM) and AlphaFold3 (AF3) through a pipeline that assesses geometric quality, energetic feasibility, fold-level dynamics, and model confidence using MolProbity, Rosetta REF2015, ProDy Gaussian Network Mode (GNM) analysis, and pLDDT, respectively. We differentiate between measurement and inference to interpret structural quality metrics for predicted structures. Our results show that the metrics differ substantially in terms of the structural properties they measure, the structural states they evaluate, and the strength of the inferences they support. For example, we found that structures with acceptable clash scores in MolProbity showed localized steric strains in the energetics analysis in Rosetta. Similarly, structures with gross backbone and side-chain errors in MolProbity had acceptable fold-level dynamics, as measured by the ProDy GNM. These findings suggest that structural quality is a multidimensional profile rather than a single score, and that confidence in predicted structures should be constructed from complementary geometric, energetic, dynamic, and interaction-based assessments while accounting for biological context, methodological limitations, and prediction uncertainty, rather than being treated as a universal indicator of the model quality.