What Does pLDDT Mean in an AlphaFold Structure?

What pLDDT is actually measuring

pLDDT stands for predicted Local Distance Difference Test. The lDDT part is an established way of scoring how close a model is to a true structure: for each residue, you look at the distances between it and its neighbouring atoms, and check how many of those distances match the real structure within a tolerance. It is a superposition-free measure, which matters — you never have to align the two structures, so a good local score cannot be spoiled by a bad global fit.

The p is the interesting part. AlphaFold does not have the true structure, so it predicts what its own lDDT score would be if you could compare it. pLDDT is therefore the model's per-residue estimate of its own local accuracy, on a 0–100 scale where higher is better.

Two consequences follow, and both are easy to miss. It is per-residue, not per-structure: every residue has its own value, and a mean over the whole chain hides a great deal. And it is local: it is a statement about whether a residue's immediate structural neighbourhood is right, and specifically not a statement about where that neighbourhood sits relative to the rest of the protein.

The four bands and the colours

AlphaFold DB uses four documented confidence bands, and the default colouring you see in any viewer maps directly onto them:

• pLDDT above 90 — very high, dark blue. The backbone and generally the side-chain orientations should be reliable. Fine for reasoning about a specific residue's local environment.
• pLDDT 70 to 90 — confident, light blue. The backbone is expected to be largely correct. Reasonable for domain-level interpretation; be more careful about individual side chains.
• pLDDT 50 to 70 — low, yellow. Treat with caution. The general fold may be in the right family but details should not be relied on.
• pLDDT below 50 — very low, orange. Should not be interpreted as structure at all. As the next-but-one section explains, this is often genuinely informative rather than simply a failure.

The local/global trap

This is the mistake worth internalising, because it produces confident and wrong conclusions.

Imagine a two-domain protein joined by a flexible linker. AlphaFold predicts both domains beautifully — pLDDT above 90 across each of them — and the linker at 40. Because pLDDT is local, both domains are legitimately reliable, and you can trust the internal geometry of each. What you cannot trust is where they are relative to each other. The model has to draw them somewhere, so it puts them in some arrangement, and that arrangement may be essentially arbitrary.

Anyone who measures a distance between a residue in one domain and a residue in the other, or reasons about a composite binding pocket formed at the interface between them, has just used the prediction to answer a question its confidence metric never addressed. The residues involved may all be dark blue. The number they produce is still unsupported.

The rule of thumb: pLDDT licenses claims about a residue and its neighbours. Any claim spanning two separately-folded regions needs a different metric.

PAE, the metric that answers the other question

That different metric is PAE, predicted aligned error, and it exists specifically for relative positioning. Where pLDDT is one number per residue, PAE is a matrix with a value for every pair of residues, reported in Ångströms.

The definition is worth reading twice: PAE at position (x, y) is the expected error in the position of residue x, if you aligned the predicted and true structures on residue y. So a low PAE between a residue in domain A and a residue in domain B means that fixing your view on one puts the other roughly where it should be — the relative placement is trustworthy. A high off-diagonal PAE means it is not, regardless of how blue both domains are.

In practice you read the PAE plot as blocks. Solid low-error squares along the diagonal are well-defined units. If two such blocks also show low error where they intersect off the diagonal, their relative orientation is supported and you can treat them as one rigid body. If the off-diagonal region is high-error, you have two confident domains whose arrangement is not determined — which is frequently the correct biological answer for a protein with a flexible linker.

Two related global scores turn up alongside these. pTM is a predicted TM-score, a single 0–1 estimate of overall fold accuracy, and ipTM is its equivalent scored across chain interfaces in multimer predictions. Both are summaries; neither replaces looking at the PAE matrix.

Low confidence is often a real result

The instinct on seeing long orange ribbons trailing off a structure is that the prediction failed there. Often the opposite is true: that is the prediction telling you something.

Regions of very low pLDDT correlate strongly with intrinsic disorder. A stretch that genuinely has no single fixed structure in solution cannot be confidently predicted to have one, and AlphaFold's low confidence is the appropriate output. Reported systematically, very low pLDDT has become a useful signal for disorder in its own right, rather than a defect to be apologised for.

The practical reading is that a long low-confidence region is a hypothesis worth checking rather than a hole in your data. Look at whether it corresponds to a known disordered region, a low-complexity segment, or a terminal tail. What you should not do is take the specific coordinates AlphaFold drew for that region and interpret them — the ribbon has to go somewhere, and where it went carries no information.

What no confidence metric tells you

Both pLDDT and PAE score how well the prediction matches a structure. Neither addresses whether that structure is the one you care about, and these gaps catch people out more often than the metrics themselves.

• Which conformational state you are looking at. Many proteins have several, and a prediction gives you one, typically without indicating which. High confidence in an apo conformation says nothing about the ligand-bound one.
• Anything about ligands, cofactors, metals or post-translational modifications. AlphaFold DB entries are the protein chain alone. A confidently-predicted active site with no cofactor in it is a confidently-predicted empty active site.
• Oligomeric state. The database predictions are single-chain, so a protein that only folds properly as a dimer is being shown out of context, and an interface surface will appear exposed.
• The effect of a mutation. Predicting the variant sequence and comparing pLDDT to the wild type is not a validated way to assess a mutation's impact — confidence tracks predictability, not stability or function.
• Whether it is your isoform. The entry corresponds to one specific UniProt accession's canonical sequence. Check the accession and the length before assuming the numbering matches your construct.

A practical reading order

Given a structure to assess, this order gets you to a defensible statement fastest.

Start with the pLDDT colouring across the whole model and identify the confident blocks and the low-confidence stretches between them — that tells you what units you have. Then, if your question spans more than one of those units, go to the PAE matrix before doing anything else, since that is where the answer about relative placement lives. Only then zoom into the specific residues you actually care about and check their individual values, rather than relying on the chain's mean.

Two habits make this concrete. Look up which residue ranges correspond to annotated domains, so you are reasoning about defined units rather than eyeballing blobs. And highlight the specific range you are interested in on the structure — a mutation site, a binding motif, a domain boundary — so you can see how its confidence compares to the model around it. SeqBench's AlphaFold Structure Viewer loads a prediction by UniProt accession with the standard pLDDT colouring already applied and lets you add labelled residue ranges in their own colours, using the same 1-based inclusive coordinates the Protein Domain Annotator reports, so domain positions paste straight across.

Related references

Related tools

Related guides

• Protein Molecular Weight and pI Calculator →
• Hydrophobicity Plots for Transmembrane Domains →
• How to Run FastQC Without Installing Anything →