Categorizing prediction modes within low-pLDDT regions of AlphaFold2 structures: near-predictive, pseudostructure and...

Area of Science:

• Structural Biology
• Computational Biology
• Protein Structure Prediction

Background:

• AlphaFold2 predictions are valuable but often contain low-confidence regions (pLDDT < 70), particularly in eukaryotic proteins.
• Interpreting these low-confidence regions is crucial for accurate structural biology applications.

Purpose of the Study:

• To identify and characterize distinct modes of behavior within low-pLDDT regions of AlphaFold2 predictions.
• To correlate these modes with known protein disorder annotations and functional elements.
• To develop a tool for users to identify and interpret these low-confidence regions.

Main Methods:

• Surveyed human proteome predictions from the AlphaFold Protein Structure Database.
• Defined and characterized three low-pLDDT behavior modes: near-predictive, barbed wire, and pseudostructure.
• Compared identified modes with disorder annotations from MobiDB and signal peptide data.

Main Results:

• Identified 'near-predictive' (potentially accurate), 'barbed wire' (no predictive value), and 'pseudostructure' (misleading elements) modes in low-pLDDT regions.
• Found correlations between barbed wire/pseudostructure and disorder, pseudostructure and signal peptides, and near-predictive regions and conditional folding.
• Developed a Phenix tool for annotating, visualizing, and selecting residues based on these prediction modes.

Conclusions:

• Characterizing low-pLDDT regions in AlphaFold2 predictions enhances their interpretability.
• The new Phenix tool aids users in understanding complex predictions and leveraging near-predictive regions for molecular replacement.