Predicted Local Distance Difference Test (pLDDT)

pLDDT provides a confidence score between 0 and 100 for each individual residue in a predicted structure. This granularity allows researchers to identify which regions of a model are reliable and which are intrinsically disordered or poorly predicted.

• High confidence (90-100): Residues are predicted with high accuracy, typically corresponding to well-ordered secondary structure elements like alpha-helices and beta-sheets
• Moderate confidence (70-90): Residues likely have a generally correct backbone but may contain local errors
• Low confidence (50-70): Regions with significant uncertainty, often corresponding to flexible loops
• Very low confidence (<50): Strong indicator of intrinsic disorder or complete lack of structural information

Unlike Root Mean Square Deviation (RMSD) or Template Modeling Score (TM-score), pLDDT does not require global superposition of structures. This makes it particularly robust for evaluating multi-domain proteins and RNA molecules where relative domain orientations may vary.

• Evaluates local atomic environments within a 15 Å inclusion radius
• Considers all atom pairs within the radius, not just Cα atoms
• Insensitive to hinge motions and domain rearrangements
• Provides a more biologically relevant assessment of model quality

In AlphaFold 2 and AlphaFold 3, pLDDT is produced by a dedicated output head trained to predict the lDDT-Cα score that the predicted structure would achieve against an experimental reference. This self-assessment is integral to the model's iterative recycling process.

• The pLDDT head is trained regressively on the difference between predicted and actual lDDT-Cα
• During inference, pLDDT values guide the recycling procedure, allowing the model to focus refinement on low-confidence regions
• AlphaFold 3 extends this to all atom types, including RNA, DNA, and ligands

pLDDT scores are used extensively to filter and interpret predicted structures in downstream applications. They provide a quantitative basis for deciding which regions of a model to trust for biological interpretation or experimental design.

• Cryo-EM model building: Low pLDDT regions often correspond to flexible domains with weak density
• Drug design: Binding site residues with high pLDDT are prioritized for docking studies
• Protein engineering: High pLDDT loops are selected as stable scaffold regions
• Intrinsic disorder prediction: Persistently low pLDDT across multiple models strongly suggests disordered regions

While pLDDT is a powerful confidence metric, it has important limitations that users must understand to avoid overinterpretation. pLDDT reflects predicted local accuracy, not global correctness or functional validity.

• High pLDDT ≠ correct fold: A confidently predicted structure can still be globally incorrect if the model is out-of-distribution
• Low pLDDT ≠ disorder: Some well-ordered regions may receive low scores if they are conformationally heterogeneous in the training set
• Not a quality of model (QMEAN) replacement: pLDDT is a self-assessment, not an independent validation metric
• RNA-specific challenges: pLDDT for RNA predictions in AlphaFold 3 is less well-calibrated than for proteins due to limited training data

The standard metric for quantifying the global similarity between a predicted 3D structure and an experimentally determined reference. RMSD calculates the average atomic distance after optimal superposition.

• Measured in Ångströms (Å); values below 2Å indicate excellent agreement
• Highly sensitive to local outliers and domain movements
• Often computed on Cα atoms only for proteins, or C3' atoms for RNA
• Used alongside TM-score for a complete structural comparison

A length-independent metric for assessing global structural similarity that is more sensitive to overall topology than RMSD. TM-score ranges from 0 to 1, with values above 0.5 indicating the same fold.

• Normalizes for protein length, enabling comparison across different-sized structures
• Less dominated by local deviations than RMSD
• A score of 1.0 indicates identical structures
• Commonly used in CASP and RNA-Puzzles benchmarks alongside pLDDT

The underlying metric from which pLDDT is derived. lDDT is a superposition-free score that evaluates local structural accuracy by comparing interatomic distances within a defined radius in the predicted model against a reference structure.

• Considers all atoms within a 15Å inclusion radius
• Invariant to rigid-body transformations, eliminating superposition artifacts
• Scores range from 0 to 100, with higher values indicating better agreement
• pLDDT is the predicted version, generated without a reference structure

A widely used metric in CASP that measures global structural similarity by calculating the percentage of Cα atoms falling within defined distance thresholds after superposition. GDT_TS is the average across four thresholds: 1, 2, 4, and 8 Å.

• Provides a balanced assessment of both local and global accuracy
• Less sensitive to flexible loop regions than RMSD
• Scores range from 0 to 100, with higher values indicating better predictions
• Often reported alongside per-residue pLDDT values for comprehensive evaluation