How well can the accuracy of comparative protein structure models be predicted?

Area of Science:

• Computational biology
• Structural bioinformatics
• Protein modeling

Background:

• Comparative protein structure models significantly outnumber experimentally determined structures.
• These models often contain errors and lack reliable accuracy assessment methods.

Purpose of the Study:

• To develop a protocol for predicting the absolute errors (Calpha root-mean-squared deviation and native overlap) of protein models without native structures.
• To enable accurate assessment of comparative model quality for practical applications.

Main Methods:

• Developed a protocol using a support vector machine (SVM) regression model.
• Constructed model-specific scoring functions using up to nine features, including sequence similarity and statistical potentials.
• Utilized tailored training sets of similar protein models based on fold or secondary structure composition.

Main Results:

• Achieved high correlation coefficients (r=0.84 for RMSD, r=0.86 for NO3.5A) in predicting model errors for over 580,000 comparative models.
• The developed scoring function outperformed 13 other assessment criteria, which showed correlations ranging from 0.35 to 0.71.

Conclusions:

• The new protocol provides an absolute error quantification for protein models, unlike traditional relative assessment scores.
• This method enhances the reliability of comparative protein models, facilitating their use in scientific research.