Predicted Aligned Error¶
An asymmetric residue-pair matrix estimating the expected positional error at one residue when a predicted protein structure is aligned on another residue's local frame.
Core Idea¶
Predicted Aligned Error is an asymmetric residue-pair matrix estimating the expected positional error at one residue when a predicted protein structure is aligned on another residue's local frame.
For residues x and y, PAE(x,y) estimates the expected positional error at residue x, measured in angstroms, if predicted and true structures were aligned on residue y's local frame. The matrix is directional: aligning on y and measuring x is not the same conditional operation as aligning on x and measuring y. Low inter-domain blocks support confidence in relative placement; high blocks warn that apparently precise packing should not be interpreted.
Scope of Application¶
The abstraction has a bounded but recurring habitat. These are literal applications of the same domain machinery, not cross-domain metaphors.
- Domain packing. between-domain blocks show confidence in relative orientation and position.
- Flexible linkers. high off-diagonal error can reveal uncertain relative motion between locally confident domains.
- Protein complexes. inter-chain blocks provide one view of relative subunit placement, alongside interface-specific metrics.
- Model triage. regions suitable for local interpretation are separated from uncertain global topology.
- Construct design. domain boundaries and flexible segments can inform experimental planning cautiously.
- Confidence visualization. heatmaps expose structure that per-residue coloring cannot.
Clarity¶
Always declare the axis convention. In the AlphaFold Database presentation, color at (x,y) is expected error at x after alignment on y. The diagonal is low by construction and is not evidence that the whole model is accurate. Interpretation belongs mainly to off-diagonal block structure.
Manages Complexity¶
PAE reduces a family of alignment-conditioned uncertainties to one matrix without collapsing them to a scalar. It separates local fold confidence from global assembly confidence and makes directionality, domains, chains, and flexible connections visible in one diagnostic surface.
The compression remains accountable because every simplification has a named validity condition. A user can ask which role is missing, which assumption fails, and which neighboring abstraction should replace the candidate instead of treating the label as an unanalyzed bundle.
Abstract Reasoning¶
R1. Record which axis is aligned and which is measured.
R2. Inspect off-diagonal blocks rather than relying on the dark diagonal.
R3. Compare PAE with pLDDT and interface metrics because they answer different questions.
R4. Treat high PAE as uncertainty, not a negative biological claim.
R5. Preserve model version and confidence calibration when comparing outputs.
Knowledge Transfer¶
PAE transfers literally among structure-prediction systems that define the same alignment-conditioned expected error. Generic pairwise uncertainty is the parent idea. Using the acronym for any symmetric contact score or coordinate RMSD drops the defining local-frame conditioning.
The transfer boundary follows from the classification test: The output recurs across AlphaFold structure assessment, while residue-axis convention, frame alignment, expected-distance bins, asymmetry, domain confidence, and model-version calibration remain constitutive. The safe portable move is to name the broader parent when the home-domain machinery is absent and to retain the domain name only when literal recognition succeeds.
Relationships to Other Abstractions¶
Current abstraction Predicted Aligned Error Domain-specific
Parents (1) — more general patterns this builds on
-
Predicted Aligned Error is a kind of Uncertainty Prime
Uncertainty. is the parent quantity retained in pairwise form.
Hierarchy path (1) — routes to 1 parentless root
- Predicted Aligned Error → Uncertainty
Neighborhood in Abstraction Space¶
Predicted Aligned Error sits in a sparse region of the domain-specific corpus (73rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (1565 abstractions)
Nearest neighbors
- Distributional Blind Spot — 0.85
- Machine-Learning Learning Curve — 0.84
- Kriging — 0.84
- Regression — 0.83
- Variational Bayesian Methods — 0.83
Computed from structural-signature embeddings · 2026-09-08