Towards an automatic classification of protein structural domains based on structural similarity

Area of Science:

• Structural bioinformatics
• Computational biology
• Protein structure analysis

Background:

• Manual protein structure classification (e.g., SCOP, CATH) is crucial for understanding evolutionary relationships but faces scalability challenges with growing datasets.
• Existing automatic methods (e.g., FSSP, Dali Domain Dictionary) produce divergent classifications, potentially due to differences in scoring metrics and procedures compared to manual methods.

Purpose of the Study:

• To investigate the impact of different hierarchical clustering procedures and pairwise similarity scores on the accuracy of automatic protein structure classification.
• To compare automatic classification results with the established SCOP fold classification and identify sources of divergence.

Main Methods:

• Utilized DALI, SHEBA, and VAST pairwise scores on SCOP C class domains.
• Applied various hierarchical clustering techniques and dendrogram-cutting strategies to generate protein structure partitions.
• Developed a method to quantify irreducible differences between automatic partitions and the SCOP classification.

Main Results:

• Ward's method clustering produced partitions most closely resembling the SCOP fold classification.
• Optimized dendrogram-cutting strategies achieved a 72% true positive rate (TPR) at a 1% false positive rate (FPR).
• Strategies not using prior SCOP knowledge, such as cutting the largest cluster, yielded competitive results (61% TPR).
• Identified substantial irreducible differences between automatic and manual classifications, indicating that global structural similarity alone is insufficient for SCOP-level classification.

Conclusions:

• The choice of clustering procedure significantly influences the agreement between automatic and manual protein structure classifications.
• Current automatic methods do not fully resolve the divergence observed when compared to manual classifications like SCOP.
• Additional information beyond global structural similarity appears necessary to achieve complete alignment between automatic and manual protein classification approaches.