AlphaFold pLDDT and PAE Scores: Is Your Model Good Enough?

Short answer: pLDDT is a per-residue confidence score from 0 to 100 stored in the B-factor column, and PAE is a residue-pair matrix showing how reliably two parts of the model are positioned relative to each other. Dock into a pocket whose residues score above 70 pLDDT, prefer above 90, and trust a multi-domain arrangement only when the inter-domain PAE is low.

The step most students skip is the one that decides whether the next three weeks of work mean anything. You fetch a model, it looks like a protein, and you push it straight into a docking box. But an AlphaFold entry is not a single quality of structure. One model can carry a crystal-grade active site and a 60-residue tail that is effectively noise, and both are written into the same file with the same apparent precision. This guide, written by the StemSkills Lab team (10+ years in sequence and structural bioinformatics, drug discovery and design, and multiscale molecular modeling), shows you how to read the two confidence outputs and turn them into a decision. It follows on from our guide to predicting a protein structure with AlphaFold and sits inside our computational biology skills roadmap.

What do pLDDT and PAE actually measure?

They answer two different questions: pLDDT asks “is this residue in the right place locally?”, and PAE asks “are these two residues in the right place relative to each other?”

AlphaFold produces three outputs for every prediction: the 3D coordinates, a per-residue confidence metric called pLDDT, and a Predicted Aligned Error matrix. The AlphaFold Protein Structure Database documentation defines pLDDT as the model’s predicted score on the lDDT-Cα metric, on a scale from 0 to 100, where higher is better.

PAE is a different animal. For a pair of residues (x, y), the value is AlphaFold’s expected position error at residue x if the predicted and true structures were aligned on residue y, reported in ångströms. The database FAQ is explicit that the matrix is asymmetrical, so the value at (x, y) need not equal the value at (y, x). In the AlphaFold DB JSON file the minimum possible value is 0 and max_predicted_aligned_error is 31.75.

This distinction matters because a two-domain protein can have every residue above 90 pLDDT and still be useless for anything that depends on the two domains sitting correctly against each other. Local confidence and global arrangement are reported separately because they fail separately.

What do the pLDDT confidence bands mean?

Four bands, and the boundaries at 90, 70 and 50 are the ones worth memorising. These come straight from the AlphaFold DB guidance, not from convention.

On the top band the documentation is unusually direct: “Regions with pLDDT > 90 are expected to be modelled to high accuracy. These should be suitable for any application that benefits from high accuracy (e.g. characterising binding sites).” That sentence is the closest thing to official permission to dock into a predicted pocket.

The bottom band carries a second meaning. The AlphaFold team showed in their paper that pLDDT below 50 is a reasonably strong predictor of disorder, meaning such a region is likely either unstructured under physiological conditions or only structured as part of a complex. A very low score there is not always a failed prediction. Sometimes it is a correct prediction that the region has no single structure to predict.

Two further rules of thumb from the same source are easy to miss. Structured domains with many inter-residue contacts are more reliable than extended linkers or isolated long helices. And unphysical bond lengths or clashes do not usually appear in confident regions, so a stretch carrying several of them should be disregarded regardless of what its score says.

Where is pLDDT stored and how do you colour a model by it?

In the B-factor column of the PDB and mmCIF files, with one trap: for a real B-factor low is good, while for pLDDT high is good. Any script that assumes “low B-factor equals well ordered” will rank your model exactly backwards.

In PyMOL, colour by the B-factor column and pin the scale to the band boundaries so that the colours mean the same thing across every model you look at:

• spectrum b, red_yellow_green, minimum=50, maximum=90

Both minimum and maximum must be supplied, otherwise PyMOL picks the range automatically from the file and two models become impossible to compare by eye. To isolate the parts you should not trust, select on the same column, for example select lowconf, b < 70 followed by show sticks, lowconf.

ChimeraX has AlphaFold support built in. alphafold fetch P00533 retrieves a model by UniProt accession and colours it by confidence by default, and adding pae true brings the error matrix along with it. The equivalent colouring command on its own is color bfactor palette alphafold, which applies the four-band scheme documented in the ChimeraX alphafold command reference: 100 to 90 high accuracy expected, 90 to 70 backbone expected to be modelled well, 70 to 50 low confidence, and 50 to 0 should not be interpreted.

Colour the model before you do anything else with it. A picture of where the confidence sits relative to your pocket answers the go/no-go question faster than any single average number.

How do you read a PAE plot?

Look for dark blocks on the diagonal, then look at the corners between them. The diagonal blocks are the confidently folded units. The off-diagonal region tells you whether those units are correctly arranged.

Read it in three steps:

1. Find the blocks. Low-error squares along the diagonal are domains. A single protein-wide low-error square means one rigid unit, which is the easiest case.
2. Read the off-diagonal corners. If the region linking two diagonal blocks is also low, AlphaFold predicts a well-defined relative position and orientation for those domains. If it is high, the documentation is clear that the relative positions and orientations “should not be interpreted”.
3. Check the stripes. A row or column of uniformly high error running the length of the plot is usually a flexible tail or linker that has no defined position against anything else.

The database FAQ adds an important caveat for anyone tempted to treat a multi-domain model as one object: in CASP14, intra-domain prediction accuracy was more extensively validated than inter-domain accuracy, and is therefore expected to be more reliable.

ChimeraX can turn this reading into an objective partition. alphafold pae uniprotId P00533 fetches the matrix, and colorDomains true clusters residues into domains using a default maximum PAE of 5.0 å between members and a minimum domain size of 10 residues. If your two “domains” survive that clustering as one group, treat them as one rigid body. If they split, treat them as two.

Is your model good enough for docking?

Judge the pocket, not the protein. The global average pLDDT printed on the entry page is the wrong number to decide on, because it is diluted by tails and linkers that your docking box will never see.

Work through this instead:

1. Identify the binding-site residues first, using a known site from the literature, a template structure, or a pocket detector. Our guide on finding a protein binding site for docking covers the options.
2. Read the pLDDT of those residues only. In PyMOL, iterate over the selection and inspect the b values directly rather than eyeballing the colours.
3. Apply the rule below, then check that the pocket sits inside one low-PAE block. A pocket formed at the interface of two domains is only as trustworthy as the inter-domain PAE that places them.

One habit will save you from most of the remaining errors: dock a ligand you already know binds, and confirm the pose reproduces the contacts reported in the literature. A model that cannot recover a known answer will not give you a trustworthy new one. Our guide on interpreting molecular docking results covers what a defensible pose looks like.

Should you trim the low-confidence loops and tails first?

Yes for docking, and almost always yes before MD. A 40-residue tail at pLDDT 30 is a prediction that the region has no fixed structure. Carrying it into a calculation gives it a single arbitrary conformation and lets it do three unhelpful things: block a pocket it never occupies, inflate your simulation box and cost, and wreck RMSD plots by dominating the fluctuation.

Trim conservatively and only at the termini and in clearly disordered loops. Cut at a point where confidence recovers, keep the structured domain intact, and never delete residues from inside a folded region simply because their scores dipped. After trimming, cap the new termini properly during system preparation, which our guide on preparing a protein for molecular dynamics walks through. Record every residue you removed in your methods notes, because a reviewer will ask.

The same logic applies in reverse for MD specifically. A model in the 70 to 90 band is a reasonable simulation starting point precisely because MD can relax an imperfect backbone, but only if you equilibrate properly and do not treat the first frame as a result. Our molecular dynamics learning path covers the equilibration protocol.

When should you use an experimental structure instead?

Whenever one exists that covers your site at reasonable resolution, and always when the chemistry you care about involves something AlphaFold does not predict.

That last point is the one that catches drug-design projects. The AlphaFold DB FAQ states plainly that AlphaFold “does not predict the positions of any non-protein components found in experimental structures (such as cofactors, metals, ligands, ions, DNA/RNA, or post-translational modifications)”. If your target is a zinc metalloenzyme, a heme protein or a kinase whose activity depends on a bound cofactor, the model gives you a pocket with a hole where the chemistry lives. The FAQ notes that side-chain geometry is often still consistent with the cofactor being present, which means you can sometimes transplant the cofactor from a homologous experimental structure, but you cannot conjure it from the prediction.

A second limitation is conformational. Where a protein is known to have multiple conformations, AlphaFold usually produces only one of them, and which one cannot be reliably controlled. For a receptor with distinct active and inactive states, you have no guarantee you received the state your project needs.

None of this makes predictions second class. In CASP14, AlphaFold reached a median GDT score of 92.4 across all targets and 87.0 in the challenging free-modelling category, against 72.8 and 61.0 for the next best methods, with a median Cα RMSD of 0.96 å to the experimental models on the best-predicted 95% of residues compared with 2.83 å for the runner-up. The database now holds 262,739,159 predicted models. For the very large number of proteins with no experimental structure at all, a high-confidence model is the only structure you are going to get. The point is to know which part of it you are allowed to use.

Whichever structure you settle on, run it through the same checks you would apply to any input, which our guide on validating a protein structure covers, and then prepare it with the standard workflow in preparing the protein and ligand for docking.

What do pLDDT and PAE not tell you?

They are confidence estimates about geometry, not about biology. The AlphaFold DB documentation is careful on this point: the scores represent AlphaFold’s confidence in the 3D arrangement of the amino acid chain, but are not a confidence in the biological correctness of the predicted model.

So a pLDDT of 95 does not tell you the protein adopts that conformation in your cell type, that the construct you will express folds the same way, or that the pocket is druggable. It tells you that if this chain has a well-defined fold, this is very likely it. Everything downstream of that, the binding, the affinity, the function, still has to be earned the usual way.

Reading confidence correctly is one of the checkpoints that separates a student project from research-grade work, and it comes early in our roadmap for building computational biology skills, well before anyone touches a trajectory.

Frequently asked questions

What is a good pLDDT score for docking?

Above 90 for the binding-site residues is the target, and above 70 is workable if you validate against a known binder. The global average for the whole chain is not the number to judge on, because tails and linkers drag it down without affecting your pocket.

Where do I find the pLDDT values in an AlphaFold PDB file?

In the B-factor column, one value per atom. Unlike a real B-factor, higher is better. Colour by it with spectrum b, red_yellow_green, minimum=50, maximum=90 in PyMOL or color bfactor palette alphafold in ChimeraX.

What does a high PAE between two domains mean?

That AlphaFold cannot place those domains relative to each other with confidence. Each domain may still be modelled well on its own. Split them and work with the relevant domain alone rather than trusting the arrangement shown in the file.

Can I run MD on a low-confidence region?

You can, but a region below 50 pLDDT starts from an arbitrary conformation, so any simulation of it explores from a meaningless origin and will dominate your RMSD and RMSF plots. Trim disordered tails before simulating, or analyse the structured domains separately.

Does a high pLDDT mean the prediction is biologically correct?

No. It is confidence in the 3D arrangement of the chain, not in biological correctness. AlphaFold also omits cofactors, metals, ligands, ions, nucleic acids and post-translational modifications, and typically returns only one of a protein’s possible conformations.

Should I use the AlphaFold DB model or run my own prediction?

The database entry is fine for most work and is instant. Run your own when your sequence is a variant, an isoform or a construct that differs from the UniProt canonical record, or when you need a complex. Note that wider UniProt entries come from a single model run rather than the best of five, which carries a slight bias toward lower confidence.

Primary sources: Jumper J. et al., Highly accurate protein structure prediction with AlphaFold, Nature 596, 583-589 (2021); and the AlphaFold Protein Structure Database FAQ maintained by EMBL-EBI and Google DeepMind, which is the reference for the confidence bands and PAE definitions quoted above.