Cryo-electron microscopy (cryo-EM) has revolutionized structural biology, leading to a massive increase in available density maps. Accurate alignment and comparison of these maps are crucial for interpreting structural information, including conformational heterogeneity analysis (via global alignment) and atomic model assembly (via local alignment). Current methods like gmfit, which uses Gaussian mixture models, and fitmap, a local optimization method in Chimera, offer varying degrees of speed and accuracy, often lacking robustness. VESPER, a vector-based approach, improves upon these methods but suffers from computational limitations due to exhaustive search. This research addresses the need for a faster and more accurate method, particularly for high-resolution maps, by introducing CryoAlign.

Existing cryo-EM map alignment methods present a trade-off between speed and accuracy. gmfit, utilizing Gaussian mixture models, provides speed but compromises accuracy, particularly with high-resolution maps. Chimera's fitmap directly optimizes voxel correlation but heavily relies on initial placement, requiring user intervention. VESPER, employing density vectors, improves accuracy and retrieval performance but is hampered by exhaustive search, making it computationally expensive. This paper aims to overcome these limitations by integrating efficient feature-based alignment with iterative refinement.

CryoAlign employs a two-stage approach. First, it generates a point cloud representation of the density map, clustering points to extract key points representing structural features. Local spatial structural feature descriptors, capturing local structural characteristics, are calculated for these key points. A mutual feature-matching strategy establishes correspondences between keypoints in different maps, enabling robust initial pose estimation. The second stage refines the alignment using a point-based iterative method, aiming to minimize distances between overlapping point pairs for optimal superimposition. The mean shift algorithm is used for key point extraction, and the DBSCAN algorithm assists in reducing the point cloud size. Density-based SHOT (signatures of histograms of orientations) feature descriptors are calculated for each key point. A two-stage alignment, initially feature-based and then point-based, refines the alignment from coarse to fine. For local alignment, a translational mask segments the larger map, treating the problem as multiple global alignment tasks. The similarity between aligned maps is measured using a function combining Jensen-Shannon divergence (for global alignment) and dot product of density vectors. Several keypoint detectors and feature descriptors are explored in the Supplementary Material.

CryoAlign significantly outperforms gmfit, fitmap, and VESPER in both global and local alignment tasks. In global alignment, CryoAlign demonstrates higher accuracy (lower RMSD) and a lower failure rate across various resolution ranges, particularly excelling with high-resolution maps. The use of key points effectively reduces computational cost while maintaining accuracy. The two-stage approach (feature-based then point-based) consistently improves alignment accuracy. In local alignment, CryoAlign's translational mask strategy addresses the challenge of size discrepancies between maps. The method's performance is robust even when the smaller map comprises less than 40% of the larger map's volume. CryoAlign also demonstrates its applicability in map comparison and atomic model fitting. In map comparison, CryoAlign facilitates the identification of variable regions during heterogeneity analysis, providing more accurate difference maps than VESPER and fitmap. In atomic model fitting, CryoAlign effectively aligns individual protein chains to the complete protein complex, outperforming existing methods, especially when dealing with rotational invariance.

CryoAlign's superior performance stems from its effective use of local spatial structural feature descriptors and a two-stage alignment process. The feature descriptors capture detailed structural information, enabling robust initial pose estimation, while the point-based refinement ensures high precision. The translational mask strategy effectively addresses the challenges of local alignment with significant size differences. The findings have implications for various cryo-EM applications, including conformational heterogeneity analysis, 3D classification, and atomic model building. The ability to accurately align maps enhances the interpretation of structural variations and accelerates atomic model assembly.

CryoAlign offers a significant advancement in cryo-EM density map alignment. Its superior accuracy and efficiency across global and local alignment tasks make it a valuable tool for researchers. Future work could focus on optimizing the segmentation strategy for local alignment and exploring different feature descriptors to further enhance performance. The algorithm is publicly available to facilitate wider adoption and further development within the cryo-EM community.

CryoAlign's current implementation focuses on maps with resolutions higher than 10 Å. The translational mask strategy in local alignment, while effective, could be improved with more sophisticated segmentation methods. The algorithm's performance might be affected by extremely low signal-to-noise ratios, limiting its applicability to certain tasks such as sub-volume alignment in subtomogram averaging. The parameter settings used in the paper might not be optimal for all tasks; users should adjust settings according to their specific needs.