Categorizing prediction modes within low-pLDDT regions of AlphaFold2 structures.

Area of Science:

• Structural biology
• Computational biology
• Protein structure prediction

Background:

• AlphaFold2 provides widely used protein structure predictions.
• Many predictions, especially for eukaryotes, include regions below the pLDDT 70 confidence threshold.
• Interpreting these low-confidence regions is crucial for accurate structural analysis.

Purpose of the Study:

• To identify and characterize distinct behavioral modes within low-confidence regions of AlphaFold2 protein structure predictions.
• To correlate these modes with known protein features like disorder and signal peptides.
• To develop a tool for users to better interpret and utilize these low-confidence prediction regions.

Main Methods:

• Analysis of human proteome predictions from the AlphaFold Protein Structure Database.
• Categorization of low-pLDDT regions into 'near-predictive', 'barbed wire', and 'pseudostructure' modes based on structural characteristics.
• Comparison of identified modes with disorder annotations from MobiDB.

Main Results:

• Three major modes were identified: 'near-predictive' (resembling folded protein), 'barbed wire' (unproteinlike, likely unpredicted), and 'pseudostructure' (intermediate, with false secondary structures).
• 'Barbed wire' and 'pseudostructure' modes generally correlate with protein disorder.
• 'Pseudostructure' is associated with signal peptides, and 'near-predictive' regions with conditional folding.
• A new Phenix tool was developed to annotate, visualize, and select residues based on these prediction modes.

Conclusions:

• Understanding low-pLDDT regions in AlphaFold2 predictions is essential for accurate structural biology.
• The identified modes provide a framework for interpreting challenging predictions.
• The developed Phenix tool aids users in distinguishing reliable from unreliable prediction regions, particularly useful for molecular replacement strategies.