A new check for cryo-EM assemblies reads only interfaces: DPI-score skips the density map and reaches AUC 0.99 on PowerFit fits
Synopsis
Model validation for cryo-EM assemblies has relied on density-fit metrics, leaving the quality of protein-protein interfaces without a direct check. DPI-score uses only the raw 3D coordinates of interface atoms, placed on a dynamically scaled voxel grid and passed to a 3D convolutional network, and reaches 87.53% validation accuracy in cross-validation. On PowerFit rigid-body fits its ROC-AUC is 0.99, above 0.46 for DeepRank-GNN-ESM and 0.54 for PI-score; on CASP15 and CASP16 the gap narrows, and on CASP16 PI-score scores higher.
Interpretation
Cryo-EM assembly validation gains a ruler aimed directly at protein-protein interfaces, one that never reads the density map and looks only at where interface atoms sit. Earlier validation metrics mostly measured how well a model matches the density map or its overall geometry, while interface metrics such as PI-score depended on derived features like contact pairs, surface area and shape complementarity, which are costly and need several tools. In five-fold cross-validation on PDB-derived interfaces, the final DPI-IV variant reached 87.53% validation accuracy; on PowerFit fits its ROC-AUC was 0.99, against 0.46 for DeepRank-GNN-ESM and 0.54 for PI-score.
The method lays interface atom coordinates straight into a 3D voxel grid and lets a 3D convolutional network judge whether that interface is close to the real structure or off. The grid scales dynamically: it first covers the interface at 1-angstrom voxels and is then rescaled to a standard size, so a fixed grid does not crush several atoms into one voxel and lose information. Training used high-resolution PDB complexes plus docked near-native interfaces and negatives far from native; testing used oligomeric cryo-EM targets from CASP15 and CASP16 and PowerFit rigid-body fits, none of which entered training.
On EMDB fitted entries worse than 3 angstroms, DPI-score flags interface errors that density-based metrics miss. After scoring 29,120 interfaces from 6,011 fitted entries, DPI-score showed very low correlation with most density-based scores, so interface quality is not simply set by how well the model fits the map. Among 6E1K fits, the misoriented 6E1K_A_B_17 received a DPI-score of 0.09 yet a CCC of 0.61, higher than the near-target 6E1K_A_B_1; re-refining 7QIN and 6UBY resolved interface clashes and improved both DPI-score and density-based metrics.
Perspective
The demonstrated scope is cryo-EM protein-protein interfaces: cross-validation on high-resolution PDB structures, oligomeric CASP15 and CASP16 targets, PowerFit rigid-body fits, and 6,011 fitted EMDB entries worse than 3 angstroms. The tool targets structural biologists building and validating cryo-EM models and is offered as Python code and as a graphical validation job in the CCP-EM Doppio suite.
On CASP15 and CASP16 the ROC curves stay above the random baseline, but AUCs are only 0.61 and 0.58, with substantial overlap between positive and negative score distributions; the source gives two different values for the CASP15 DPI-score AUC, 0.61 in one place and 0.65 in another, and they do not agree. On CASP16 PI-score reached an AUC of 0.78 against 0.58 for DPI-score. Many interfaces in all three test sets fell outside the IS-score and iRMSD thresholds and were left unlabelled, about 92% for CASP15, 43% for CASP16 and 88% for PowerFit, and negative labels are rare. For targets such as H1137, T1234o and H1236, interfaces with low IS-score but iRMSD of at most 3 angstroms received high DPI-scores, and the authors suggest atom positions alone may not separate such near-native models.
