Reading Confidence Scores

Why confidence scores matter

AlphaFold will always produce a prediction, even for regions where the model has no real idea what the structure should be. The model doesn't "refuse" to make a prediction for disordered or poorly constrained regions — it always outputs coordinates. Without reading the confidence scores, you cannot know whether to trust a prediction.

This is one of the most important practical lessons in this guide. Many beginners look at an AlphaFold structure, see a beautifully coloured ribbon diagram, and assume it's all reliable. It isn't. Some regions are predicted with atomic-level accuracy; others are essentially guesses that happen to look structured.

AlphaFold provides two main confidence measures:

1. pLDDT — per-residue local confidence (how well is each individual position predicted?)
2. PAE — pairwise confidence (how certain is AlphaFold about the relative positions of two parts of the protein?)

pLDDT: predicted local distance difference test

pLDDT stands for predicted local distance difference test. Let's break that down.

The local distance difference test (LDDT) is a real quality metric used to compare experimental and predicted structures. For a given residue, it measures how well the predicted distances to all neighbouring atoms match the true distances. Specifically, it asks: "For all atom pairs within 15 Å of this residue, what fraction of predicted distances are within a tolerance threshold (0.5 Å, 1 Å, 2 Å, 4 Å) of the true distances?"

The predicted LDDT (pLDDT) is AlphaFold's estimate of what this score would be for each residue, before you have the experimental structure to compare against. It's the model's self-assessment of its own confidence.

pLDDT is expressed as a score from 0 to 100, where 100 means the model is extremely confident its predicted local geometry is correct.

The pLDDT colour scale

The standard AlphaFold colour scale encodes pLDDT as follows. Crucially, this is a non-linear scale — the colours don't divide the 0–100 range evenly, they're chosen to be meaningful for interpretation:

Interpreting each colour in practice

The dark blue residues at the top are in a well-packed protein core — AlphaFold is highly confident about their positions. The orange Ser at the bottom is in a flexible loop that may be intrinsically disordered in the real protein.

PAE: predicted aligned error

pLDDT tells you about local confidence — how well is each residue predicted in isolation? But proteins often have multiple domains, and pLDDT can't tell you whether AlphaFold is confident about their relative positions and orientations.

For this, AlphaFold provides the Predicted Aligned Error (PAE) matrix. PAE is an N×N matrix where each entry PAE(i, j) gives AlphaFold's estimate of the expected error (in Ångströms) in the position of residue i's Cα atom, when the structures are aligned on residue j.

In plainer terms:

• Low PAE(i, j): AlphaFold is confident about the relative position of residues i and j.
• High PAE(i, j): AlphaFold is uncertain — residues i and j might be in quite different relative positions in the real structure.

Reading a PAE heatmap

PAE is visualised as a heatmap where dark colours (typically dark green or dark blue depending on the viewer) indicate low error (confident) and light colours indicate high error (uncertain).

The axes of the PAE plot both run from residue 1 to residue N. The diagonal always has very low values — trivially, a residue's position relative to itself is perfectly certain.

Here's an idealised diagram of what different PAE patterns mean:

Contrast this with a single compact domain:

When studying multi-domain proteins or protein complexes, always check the PAE. A protein can have all dark-blue pLDDT but a highly uncertain PAE, meaning: "each domain is well-structured, but they can wobble relative to each other." This matters enormously for understanding protein function and for modelling interactions.

pTM and ipTM: overall quality scores

Beyond pLDDT and PAE, AlphaFold provides two summary confidence scores:

• pTM (predicted TM-score): An estimate of the TM-score — a global structure similarity metric — that the predicted structure would achieve against the true structure. TM-score > 0.5 generally indicates a correct fold. pTM > 0.7 is considered a reliable prediction.
• ipTM (interface pTM): For complex predictions (AlphaFold-Multimer or AlphaFold 3), ipTM specifically measures confidence in the predicted protein–protein interface. ipTM > 0.8 is considered a confident complex prediction. A combined metric 0.8 × ipTM + 0.2 × pTM is often used to rank complex predictions.

Common confidence interpretation mistakes

• Ignoring confidence entirely. Looking at the structure without reading pLDDT or PAE. Low-confidence regions can look just as "pretty" as high-confidence ones in default visualisations.
• Trusting orange/yellow loops as real structure. Low-pLDDT regions often correspond to intrinsically disordered regions. Treating them as structured can lead to wrong biological conclusions.
• Not checking PAE for multi-domain proteins. Two domains can each be individually high-confidence (dark blue pLDDT) while their relative orientation is completely uncertain (high PAE). Without the PAE, you miss this.
• Assuming high pLDDT means the exact conformation is the only one. Proteins are flexible and can adopt multiple conformations. AlphaFold gives one conformation (or a few with AF3). Even high-confidence predictions may not represent the biologically active conformation in all conditions.
• Treating a pLDDT of 72 as dramatically different from 71. Confidence scores have uncertainty. The boundaries (50, 70, 90) are useful heuristics, not precise thresholds. A residue at 68 and one at 72 are essentially in the same confidence tier.