Predicted Local Distance Difference Test (pLDDT)

pLDDT is a per-residue confidence score ranging from 0 to 100. It estimates the local distance difference test (lDDT) score that would be achieved against an experimental reference structure, without requiring that reference. A higher score indicates higher predicted accuracy for that specific residue's local environment.

• Scale: 0 (low confidence) to 100 (high confidence)
• Granularity: Computed independently for each residue
• Interpretation: A proxy for the expected lDDT-Cα value

AlphaFold categorizes residues into discrete confidence bands for intuitive interpretation and visualization:

• Very high (pLDDT > 90): High accuracy; suitable for detailed structural analysis and characterizing binding sites.
• Confident (90 > pLDDT > 70): Generally good backbone prediction; useful for domain-level analysis.
• Low (70 > pLDDT > 50): Low confidence; treat with caution, often corresponds to flexible loops.
• Very low (pLDDT < 50): Strong indicator of intrinsic disorder; the residue likely does not adopt a single stable conformation in isolation.

pLDDT is often confused with the Predicted Aligned Error (PAE), but they measure fundamentally different aspects of confidence:

• pLDDT: Measures local confidence. It answers: "How accurate is the local packing around this residue?"
• PAE: Measures global confidence. It answers: "If I align the prediction on residue i, how far off is the predicted position of residue j?"
• Key Insight: A domain can have high pLDDT (well-folded internally) but high PAE relative to another domain (uncertain relative orientation).

The pLDDT predictor is a regression head within the AlphaFold architecture, trained in a supervised manner:

• Training Signal: The actual lDDT-Cα score, calculated by comparing a predicted structure to its known experimental structure in the Protein Data Bank (PDB).
• Self-Consistency: During training, the model learns to estimate the lDDT it would achieve, making it a self-assessment mechanism.
• Output: The final pLDDT value is a scalar per residue, extracted from the model's output before the final structure is generated.

pLDDT is the primary gatekeeper for downstream computational and experimental analyses:

• Disorder Prediction: Residues with pLDDT < 50 are a state-of-the-art predictor of intrinsically disordered regions (IDRs).
• Domain Parsing: High-confidence regions separated by low-confidence linkers delineate structured domains.
• Variant Interpretation: The impact of a missense mutation is more reliably modeled in high-pLDDT regions.
• Molecular Docking: Only high-confidence (pLDDT > 90) side-chain positions should be used for rigid-body docking studies.

Despite its utility, pLDDT has critical limitations that must be understood to avoid over-interpretation:

• Not a Global Score: A high average pLDDT does not guarantee correct domain orientation; always cross-reference with PAE.
• Confidence ≠ Function: A high pLDDT score confirms structural accuracy but does not validate a functional hypothesis.
• Artifactual Folding: In rare cases, a sequence may be predicted with high confidence in a non-native conformation, especially in the absence of deep MSAs.
• Ligand Blindness: pLDDT does not reflect confidence in the position of co-factors, ligands, or ions.