PAE (Predicted Aligned Error)

PAE provides a distance error estimate in Ångströms for every pair of residues (i, j) in a predicted structure. The output is a 2D heatmap where the value at position (i, j) represents AlphaFold's expected error in the position of residue i if the predicted and true structures were aligned on residue j.

• Low PAE (< 5 Å): High confidence in the relative position of the two residues
• High PAE (> 15 Å): Low confidence; the relative orientation is uncertain
• The matrix is asymmetric—PAE(i,j) is not necessarily equal to PAE(j,i)

PAE is the primary diagnostic tool for evaluating inter-domain accuracy in multi-domain proteins. While pLDDT reports per-residue confidence, it does not reveal whether two well-predicted domains are correctly oriented relative to each other.

• Low inter-domain PAE: Indicates the model is confident about the relative orientation and packing of the two domains
• High inter-domain PAE: Suggests the domains may be folded correctly individually but their relative position is uncertain
• Look for distinct PAE blocks along the diagonal to identify domain boundaries

The PAE plot is a color-coded 2D matrix where both axes represent the residue index. Dark green regions indicate low predicted error, while white or light regions indicate high uncertainty.

• Diagonal blocks: Well-predicted domains appear as dark green squares along the diagonal
• Off-diagonal signals: The color between two diagonal blocks reveals the confidence in their relative orientation
• Deep green off-diagonal: The two domains have a confidently predicted relative pose
• White or pale off-diagonal: The relative domain orientation is essentially unconstrained by the model

PAE and pLDDT (predicted Local Distance Difference Test) serve complementary roles in structure quality assessment. Understanding their distinction is critical for proper model evaluation.

• pLDDT: Per-residue confidence in local backbone geometry; high pLDDT means the local fold is correct
• PAE: Pairwise confidence in relative positions; low PAE means the spatial relationship between two residues is reliable
• A residue can have high pLDDT but high PAE relative to residues in another domain—the local structure is correct, but the domain's position in space is uncertain
• Always examine both metrics before drawing biological conclusions from a predicted structure

PAE is produced by a dedicated structure module head within AlphaFold2's architecture. During training, the model learns to predict the error it would make when aligning the predicted structure to the ground truth.

• The PAE head outputs a probability distribution over distance bins for each residue pair
• The reported PAE value is the expected distance error computed from this distribution
• This mechanism is trained end-to-end with the rest of the network, providing a calibrated uncertainty estimate rather than a post-hoc heuristic
• The recycling mechanism iteratively refines both the structure and the PAE estimates across multiple passes

PAE guides critical decisions in structural biology workflows and computational pipelines:

• Domain segmentation: Use PAE matrices to automatically identify structurally independent domains for modular analysis
• Docking validation: Assess whether predicted protein-protein interfaces have confident relative orientations before investing in experimental validation
• Cryo-EM model building: Cross-reference PAE with experimental density maps to identify regions where the prediction is reliable enough to guide manual model building
• Filtering predictions: In high-throughput structure prediction projects, PAE thresholds can automatically flag low-confidence models for exclusion or further refinement