Using deep-learning predictions reveals a large number of register errors in PDB depositions

Area of Science:

• Structural Biology
• Computational Biology
• Biochemistry

Background:

• Protein Data Bank (PDB) accuracy is crucial for downstream applications.
• Experimental data limitations, especially at low resolution, can introduce errors.
• Existing validation methods (stereochemistry, map-model agreement) have limitations.

Purpose of the Study:

• To introduce and evaluate a novel, resolution-independent protein structure validation method.
• To identify and correct register errors in PDB structures.
• To improve the overall accuracy and reliability of protein structural data.

Main Methods:

• Developed a validation approach comparing observed residue contacts/distances with AlphaFold2 predictions.
• Applied the method to scan 3-5 Å resolution structures in the PDB.
• Implemented suggested corrections and assessed their impact on refinement statistics.

Main Results:

• Identified thousands of likely register errors in PDB structures.
• Demonstrated that suggested corrections improve refinement statistics in most cases.
• Characterized limitations such as fold-switching proteins.

Conclusions:

• The novel validation approach is effective in detecting register errors, orthogonal to traditional methods, and resolution-independent.
• Implementation of suggested corrections enhances structural model quality.
• Expected to improve accuracy of current and future PDB depositions via CCP4 integration.