Skip to content

Reading Confidence Metrics: pLDDT, PAE, pTM & ipTM

AlphaFold and ColabFold don't just give you a structure — they give you an estimate of how much to trust each part of it. Learning to read that estimate is the single most important skill for using these tools responsibly. The short version: pLDDT tells you about one residue; PAE tells you about the relationship between two residues; pTM/ipTM boil an entire structure or interface down to one number. None of them substitutes for the others.

See it on a real structure first

Before the theory, Try It: A Real Example Protein puts an interactive 3D viewer with live pLDDT coloring in front of you — worth a look before or after reading this page.

Try it: what does your score mean?

Drag the sliders to your own numbers and get a plain-language read — the same thresholds explained below, applied live.

pLDDT — per-residue local confidence

Beginner

pLDDT (predicted Local Distance Difference Test) is a 0–100 score for each residue, estimating how well that residue's local environment matches reality. Higher is better. It's usually shown as the coloring on the 3D cartoon:

Range Label What it means
> 90 Very high Backbone and side chains placed with high precision
70–90 Confident Backbone generally right; side chains may be off
50–70 Low Uncertain — interpret with caution
< 50 Very low Probably wrong; often disordered

Horizontal pLDDT confidence scale from 0 to 100, colored orange below 50, yellow 50-70, light blue 70-90, dark blue above 90

Low-pLDDT stretches are frequently intrinsically disordered regions (IDRs) — parts of the protein with no single fixed structure — or simply regions where AlphaFold didn't have enough evolutionary information to work with (e.g. flexible linkers between domains). See these exact bands on a real protein in Try It: A Real Example Protein — 80% of BRCA1 is "very low," and that's not a failed prediction.

Advanced

pLDDT is computed from four binary distance-agreement tests at increasingly strict thresholds (4 Å, 2 Å, 1 Å, 0.5 Å) between predicted and (during training) true inter-atomic distances; the reported score is essentially the average pass rate across those thresholds. In AlphaFold3, pLDDT becomes a per-atom score computed for every molecule type (proteins, nucleic acids, ligands) rather than only per-residue for proteins.

Practical use: read the ChopChopMF pLDDT coloring / AlphaSync workflow to get this per-residue table directly in ChimeraX rather than eyeballing colors.

Expert: pLDDT is not a clean disorder detector

Several failure modes are easy to miss:

  • A low-pLDDT helix usually means a transient helix, not a wrong one. Because AlphaFold outputs a single fixed set of coordinates, it cannot directly represent a conformational ensemble. When a local sequence has some helix-forming propensity (e.g. an amphipathic pattern, or a short Molecular Recognition Feature/MoRF that folds only upon binding a partner) but sits within an otherwise intrinsically disordered region (IDR), the model will often still draw a helix — but flags it with pLDDT in the 50–70 range or below. Read that combination (helical geometry + low pLDDT) as "this helix probably forms only transiently or conditionally," not as a modeling error to discard. One study comparing AF2 predictions to NMR/SAXS conformational ensembles found pLDDT and structural heterogeneity across the ensemble "agreed closely with the observed secondary structure from the ensembles" — i.e. a lower-pLDDT helix among two candidate helices is consistent with it existing only transiently, while conversely a coiled region with unexpectedly high pLDDT can itself be a signal of a disorder-to-order transition or a conserved role in some biophysical interaction (PMC9104326).
  • Conditionally-folded IDRs can score deceptively high, not just low. Some IDRs only adopt a defined structure when bound to a partner or after a post-translational modification. AlphaFold, given only the naked sequence, can still predict these with high pLDDT — apparently because the co-evolutionary signal implicitly encodes the bound/modified conformational state. A high-confidence prediction for a region known (or suspected) to be an IDR should not automatically be read as "stably folded in isolation." Researchers have used exactly this signature to systematically hunt for conditionally-folding IDRs at proteome scale (PMC10622901).
  • ~10% of low-pLDDT stretches aren't about disorder at all — other sources of local uncertainty (shallow MSA, genuinely hard local geometry) can also depress the score. Don't treat pLDDT < 50 as synonymous with "disordered" without corroborating a dedicated disorder predictor or experimental evidence.
  • A minor, preliminary observation worth knowing about: one community-reported analysis found that the geometric angle between a residue's backbone-propagation direction and its hydrogen-bond vector in secondary structure elements — a near-constant ~23° for alpha-helices and ~13° for antiparallel beta-sheets in real structures — deviates more in low-pLDDT regions than high-pLDDT ones, with low-confidence helices showing roughly 79% more severe geometric violations (GitHub discussion). The reported effect size is tiny (~0.7% of variance explained), so this is at most a supplementary sanity signal, not a validated tool — but it's a reminder that pLDDT correlates with more than just "is this residue in the right place": subtle backbone geometry quality tracks with it too.

AlphaFold3-specific: spurious/hallucinated helices

AlphaFold2 is non-generative and typically renders disordered regions as loose, unstructured "spaghetti" ribbons — visually easy to recognize as low-confidence. AlphaFold3 is a generative (diffusion-based) model, and per EBI's training material it "sometimes predicts spurious structural order (hallucinations) in disordered regions" — including forming alpha-helical structures instead of the expected ribbon-like appearance. These hallucinated helices are still "usually marked as very low confidence, with pLDDT scores well below 50," but they lack AF2's telltale unstructured look, so in AF3 output, low pLDDT is the primary — sometimes the only — signal that a confident-looking helix isn't real. EBI notes such hallucinations are rare overall, but the practical implication is: in AF3 models, never let a clean-looking helix override a low pLDDT score. (EBI: What AlphaFold 3 struggles with)

PAE — Predicted Aligned Error

Beginner

In plain terms: PAE is how sure AlphaFold is that two residues are positioned correctly relative to each other in space — not "is this residue folded right" (that's pLDDT), but "if residue y is exactly where I put it, how far off would residue x be?" It's measured for every pair of residues, shown as a 2D heatmap (both axes = residue number), usually with dark = low error/confident, light = high error/unreliable.

The diagonal is always dark (a residue compared to itself) — ignore it. What matters are the off-diagonal blocks: a dark block between residue ranges of two domains (or two chains in a complex) means AlphaFold is confident about how they're positioned relative to each other. A light block means that relative placement is essentially a guess.

Schematic comparison of two PAE heatmaps: one with a dark, confident off-diagonal block between two domains, one with a light, uncertain off-diagonal block despite both domains individually being well-folded

For what this looks like on an actual entry — a single small confident block against an otherwise uniformly light background — see the real PAE plot for BRCA1.

Why do I need PAE if I already have pLDDT?

Because they answer genuinely different questions, and pLDDT structurally cannot answer the one PAE answers. pLDDT is purely local — it's telling you "is residue N's own immediate environment modeled correctly," one residue at a time. It has no mechanism for expressing "domain A and domain B are each folded perfectly, but I have no idea how they're rotated relative to each other" — and that exact situation is common. A structure can show uniformly high pLDDT in two well-folded domains while PAE reveals their relative orientation is essentially random. This is the single most common way people over-trust a multi-domain or multimeric prediction: they see a confident-blue cartoon, and stop looking. Any claim about how two domains or two chains sit relative to each other needs the corresponding off-diagonal PAE block — pLDDT alone cannot tell you.

Advanced: chance or determination? — why PAE is what makes multimer predictions useful

Reframing the question this way is the key intuition for complexes: for a residue pair spanning two different chains, low PAE means "AlphaFold has real, determined evidence for exactly this relative placement," while high PAE means "AlphaFold has no meaningful basis for this arrangement — treat the two chains' relative pose as essentially undetermined."

This distinction matters enormously in practice because of a specific, well-documented behavior: when you give AlphaFold/ColabFold two chains to predict as a complex, it will place them near each other in the output regardless of whether they actually interact — there is no "these don't bind" output option; the model must produce some single relative pose. So visually seeing two chains in contact in the 3D cartoon proves nothing on its own — that contact could be a genuine, confidently-determined interface, or it could be an artifact of the model having to put the second chain somewhere. The inter-chain PAE block is what tells the two cases apart: low PAE across the interface residues means the proximity you're looking at is a determined prediction; high PAE means it's closer to arbitrary, and you should not read biological meaning into that contact without independent evidence.

For complexes, the inter-chain blocks of the PAE matrix are therefore your primary evidence for whether a predicted interface is real — this is exactly what the ipTM score (next section) tries to summarize into one number. ChopChopMF's PAE Contacts workflow turns this into visible pseudobonds in ChimeraX; for a full all-vs-all heatmap across a large complex, the standalone PAE Viewer is a useful companion.

Expert: why Ångström, precisely — and related metrics

PAE is literally a 3D distance-error estimate, which is why it's reported in Ångströms rather than as a unitless score like pLDDT or pTM. Mechanically: take the predicted backbone frame (N, Cα, C atoms) of residue y, superimpose/rotate it exactly onto residue y's position in the (hypothetical) true structure, and ask — with that one point pinned down as a reference frame — how far, in Å, does residue x's predicted Cα then land from where it truly is? This is discretized into 64 bins and repeated for every (x, y) pair to build the full matrix. AlphaFold3 generalizes the same idea to backbone frames for polymers and per-atom neighbor frames for ligands/PTMs, and adds a related metric, PDE (Predicted Distance Error) — pairwise distance accuracy independent of frame alignment.

Because PAE depends on which residue you anchor on, it is not symmetric in general: the error at x when aligning on y isn't necessarily the same as the error at y when aligning on x — anchoring on a residue in a large, rigid domain tends to produce lower error elsewhere in that same domain than anchoring on a residue in a small, wobbly appendage, since a small/flexible anchor is itself a less stable reference frame.

You'll sometimes see papers report "iPAE" (interface PAE) as a single summary number rather than the full matrix — this isn't an official AlphaFold output field, and users have pointed out the AlphaFold-Multimer paper doesn't spell out how to compute it. In practice it's derived post-hoc from the raw PAE matrix in the output pickle/JSON: extract the inter-chain PAE values (typically restricted to residue pairs within some distance cutoff, e.g. ~3.5–8 Å in published pipelines) and summarize them — commonly as a mean or as the lowest peak of the inter-chain PAE distribution. Because there's no single standardized definition, treat a reported "iPAE" number cautiously and check the specific paper's/tool's methods for exactly how it was computed before comparing it across studies.

If you specifically want the domain-packing/interface consistency issue solved for you with an established metric instead of a home-grown "iPAE," see actifpTM — a refined AlphaFold2 confidence metric published specifically to handle flexible/multi-domain regions better than raw ipTM.

pTM and ipTM — one number for a whole structure or interface

Beginner

  • pTM (predicted TM-score) summarizes the accuracy of the entire predicted fold in a single number, roughly 0–1. As a rule of thumb: > 0.5 means the overall fold might resemble the truth; < 0.5 means it likely doesn't.
  • ipTM (interface predicted TM-score) is the same idea, but restricted to the interface between chains in a complex — the single most useful number for "is this predicted protein-protein interaction believable?"
ipTM Interpretation
> 0.8 High-quality, confident interface
0.6 – 0.8 Gray zone — could go either way, look closer
< 0.6 Likely unreliable

Horizontal ipTM confidence scale from 0 to 1, colored red below 0.6, yellow 0.6-0.8, green above 0.8

BRCA1 on the example page is a single chain, not a complex — which is exactly why it has no ipTM at all: the score only exists for multi-chain predictions.

Advanced: don't stop at the single number

  • pTM can be dominated by a large subunit in complexes with very unequal chain sizes, masking a poorly predicted small partner — always sanity check the small chain's own pLDDT/PAE too.
  • Legitimate disorder can drag ipTM down even when the actual interface of interest is well predicted, since disordered regions elsewhere in the complex still count against the aggregate score. If ipTM looks mediocre, check whether the specific interface you care about has a dark PAE block before writing the whole prediction off.
  • AlphaFold3 reports a composite ranking_score = 0.8 × ipTM + 0.2 × pTM + 0.5 × fraction_disordered − 100 × has_clash to rank multiple output models — has_clash is a hard penalty for physically implausible structures (heavy atomic clashes).

Real-world example: why rank-0 alone isn't enough

A user screening ligand-receptor pairs with AlphaFold-Multimer 2.3.1 reported ipTM scores ranging from 0.30 to 0.88 across the five models for the same input pair — and asked whether to trust only the top model or average across all five. There's no universal answer, but the guidance above gives you a way to reason about it: don't just take the single highest ipTM at face value. Look at why the models disagree — is it a genuinely ambiguous/flexible interface (in which case the spread itself is informative — a real signal of uncertainty, not noise to average away), or is one model an outlier while four cluster together (in which case the cluster is probably more trustworthy than any single extreme)? Either way, large variance across the five models is itself a finding worth reporting, not a nuisance to collapse into one number.

Expert: known failure regimes
  • For structures/complexes under roughly 20 tokens, pTM becomes statistically unreliable (pinned near ~0.05) regardless of actual quality — trust PAE and pLDDT instead in that regime.
  • Some large-scale interaction-screening pipelines (using speed-optimized, minimal-recycling settings) have used ipTM cutoffs as low as ~0.3 purely as a first-pass filter over huge candidate sets — but hits at that threshold still require independent, targeted verification and should never be treated as confident findings on their own.
  • Confidence scores in AlphaFold3/Server are context-dependent, not a fixed property of the sequence: adding or removing ions, cofactors, ligands, or other chains from the same job changes the reported pLDDT/PAE/ipTM for a given protein. Always evaluate a prediction under the biologically relevant input composition, not a stripped-down convenience version of it.

A practical trust-checklist

Advanced workflow

  1. Start with pLDDT to triage which regions are worth looking at closely (≥70) vs. treating as noise/disorder (<50).
  2. Check the PAE off-diagonal blocks for every domain-domain or chain-chain relationship your conclusion depends on — never infer relative positioning from pLDDT alone.
  3. For complexes, read pTM + ipTM together, but cross-reference the specific interface's PAE block before trusting or dismissing based on the aggregate score alone.
  4. Sanity-check low-confidence regions against biology: is disorder expected there (terminus, known flexible linker)? If disorder would be biologically surprising, treat the low score as a possible prediction failure rather than true flexibility.
  5. Don't rely on rank-0 alone — AlphaFold/ColabFold typically output 5 ranked models; check whether your conclusion holds across the top models, not just the single best-ranked one.
  6. Recommended reading order for a fresh result: MSA/sequence-coverage plot → pLDDT plot → PAE plot → 3D structure. Forming an expectation of reliability before looking at the cartoon avoids being visually seduced by a nice-looking but poorly-supported fold. See Presenting & Sharing Results for how to report all of this in a paper.
flowchart TD
    A[New prediction] --> B{pLDDT per region}
    B -->|"< 50"| C[Treat as noise / likely disordered]
    B -->|"50-70"| D[Uncertain - flag, don't build conclusions on it alone]
    B -->|"&ge; 70"| E{Relationship between domains/chains matters?}
    E -->|No| F[Reasonably trustworthy on its own]
    E -->|Yes| G[Check PAE off-diagonal block]
    G -->|Light/high error| H[Relative position is a guess - do not trust]
    G -->|Dark/low error| I{Multi-chain complex?}
    I -->|No| F
    I -->|Yes| J[Check pTM + ipTM]
    J -->|"ipTM > 0.8"| F
    J -->|"0.6-0.8"| K[Gray zone - corroborate with biology / other evidence]
    J -->|"< 0.6"| H

Worked example: running this checklist on BRCA1

Applying it to the protein from Try It: A Real Example Protein: ~80% of BRCA1 sits below pLDDT 50, so step 1 triages most of it as noise/likely-disordered — matching what's independently known about BRCA1's large IDRs. The one region that clears 70, the C-terminal BRCT domain, is exactly the single dark block in its real PAE plot, so its local fold is trustworthy on its own (step 2). It's a monomer, so steps 3 and 5 (ipTM, cross-model agreement for a complex) simply don't apply here — there's no interface to check.

One more input factor is worth knowing before you trust — or dismiss — a score: see What Actually Influences a Prediction? for why the sequence/MSA matters far more than you might expect, and why a single point mutation rarely moves these numbers the way you'd hope.

Continue to What Actually Influences a Prediction? →