Using deep-learning predictions of inter-residue distances for model validation.

Area of Science:

• Structural biology
• Computational biology
• Deep learning applications

Background:

• Protein structure determination relies on building models that fit experimental data and depositing them in the Protein Data Bank.
• Model building can introduce errors due to experimental limitations, and current validation metrics often focus on physico-chemical properties or data fit.
• Deep learning advancements have improved inter-residue distance predictions, significantly aiding protein ab initio modeling.

Purpose of the Study:

• To introduce novel protein model validation methods utilizing deep learning-based inter-residue distance predictions.
• To compare predicted distances with those observed in protein models for enhanced validation.
• To specifically address the detection and correction of sequence-register errors in protein models.

Main Methods:

• Employing deep learning algorithms for accurate prediction of inter-residue distances.
• Comparing predicted inter-residue distances with distances derived from protein structural models.
• Developing algorithms to identify sequence-register errors and determine necessary correction shifts.

Main Results:

• The new validation methods demonstrate high precision in detecting sequence-register errors within protein models.
• The proposed approach reliably determines the register shifts needed to correct identified sequence errors.
• Validation metrics based on predicted inter-residue distances offer a powerful complement to existing methods.

Conclusions:

• Deep learning-based inter-residue distance predictions offer a robust new avenue for protein model validation.
• These methods significantly improve the accuracy of error detection, particularly for sequence-register issues.
• The ConKit package provides access to these advanced validation tools for the structural biology community.