Metrics for the Evaluation of 3D Segmentation in Biomolecules

Accurately segmenting biomolecular structures into structural domains is crucial for functional and evolutionary analysis, as well as structure prediction. Over the past decades, a wide range of 3D segmentation algorithms have been developed for proteins and RNA molecules. However, the metrics for assessment of segmentation quality remain inconsistent across studies, hindering fair comparisons between methods. These metrics, with their different characteristics, may not be correlated and may yield different conclusions for the same method. In this study, we present a comprehensive overview and systematic analysis of metrics used to evaluate 3D biomolecule domain segmentation. We also summarize and categorize the widely used evaluation measures into two principal groups: domain overlap-based metrics, which quantify residue-level agreement between corresponding domains, and distance-based metrics, which focus on the proximity of predicted versus target boundaries. On each metric, we describe how it works, thereby assessing its strengths and limitations. This would be helpful in clarifying the ambiguities resulting from metric inconsistency, providing guidance for future method development, and standardizing the assessment of 3D structure parsing to ensure robustness and meaningful evaluation.