How to Read AlphaFold Confidence Scores (pLDDT and PAE Explained) (Video)
Written guide: AlphaFold Output Files Explained
This video shows how to read AlphaFold confidence scores, pLDDT and PAE, and which parts of a predicted protein model to trust.
Chapters
- 0:00 AlphaFold confidence scores
- 0:53 Where AlphaFold came from
- 2:07 What pLDDT measures
- 3:35 The four confidence bands
- 4:34 A real AlphaFold model
- 5:23 When the scores mislead
- 6:37 Predicted aligned error
- 8:01 Scores for protein complexes
- 9:13 Checking a model yourself
- 10:03 Review and question
Video transcript
0:00 AlphaFold confidence scores
In this lesson, we will learn how to read the confidence scores that AlphaFold reports, pLDDT and PAE, and what they tell us about a predicted protein structure. AlphaFold predicts the three-dimensional structure of a protein from its amino acid sequence, and it reports confidence scores alongside each predicted structure. First, we will look at where AlphaFold came from, and then we will see what the score called pLDDT measures and how to read its four color bands. With that foundation, we will read the scores of a real AlphaFold model and look at the cases in which the scores mislead. In the final sections, we will cover the predicted aligned error and the scores for protein complexes, and we will finish with ways to check a model in your own work. To begin, we need to see where AlphaFold came from and why its predictions need confidence scores at all.
0:53 Where AlphaFold came from
AlphaFold addresses a central problem in biochemistry, which is how the sequence of a protein determines its shape. A protein is a chain of amino acids, also called residues, and the chain folds into a three-dimensional shape that determines what the protein does. Far more protein sequences are known than structures, so researchers have long tried to predict a structure from the sequence alone. Since 1994, a competition called CASP has tested prediction methods in rounds held two years apart on structures that are solved but not yet released. In 2020, at the fourteenth round of CASP, AlphaFold2 from the company DeepMind predicted many structures with an accuracy close to that of experimental methods. In 2021, DeepMind and the European Bioinformatics Institute opened the AlphaFold Protein Structure Database, which has held more than 200 million predicted structures since 2022. Researchers now use these predicted structures without an experiment behind them, so they need a way to judge which regions of a model they can trust. AlphaFold supplies that judgment as confidence scores, and the first score we need is pLDDT.
2:07 What pLDDT measures
Second, we turn to the main confidence score, which AlphaFold reports for each residue of a model. The score is called the predicted local distance difference test, or pLDDT, and it ranges from 0 to 100. A higher score means that AlphaFold is more confident about the local structure around that residue. The AlphaFold Database colors each residue of a model by this score, so the confidence appears directly on the structure. To see what the score predicts, we need the measure behind it, the local distance difference test, or lDDT. The lDDT compares a model with an experimentally determined structure of the same protein. Take a single residue in the experimental structure and look at its distances to the atoms around it, out to 15 angstroms, where an angstrom is a tenth of a nanometer. The lDDT checks what fraction of those distances the model reproduces within four tolerances, 0.5, 1, 2 and 4 angstroms, and averages the four fractions. A protein can contain several domains, compact regions of the chain that fold largely on their own. Because the lDDT compares only local distances, a domain with the right fold but the wrong angle to the rest of the protein still scores high for most of its residues. When AlphaFold makes a prediction, it has no experimental structure to compare against, so pLDDT is its own estimate of the lDDT that the model would score.
3:35 The four confidence bands
Third, with the meaning of the score in place, we turn to the four bands that the AlphaFold Database uses to color pLDDT. Above 90 is the very high band, in dark blue, where AlphaFold usually places the backbone and most side chains accurately. Between 70 and 90 lies the high band, in light blue, where the backbone is generally correct but some side chains may sit in the wrong position. The low band runs from 50 to 70 and appears in yellow, and here even the backbone may be wrong, so treat the predicted positions of these residues with caution. Below 50 is the very low band, in orange. Some stretches of a protein chain, called intrinsically disordered regions, have no single stable structure on their own. The AlphaFold Database describes a score below 50 as a reasonably strong predictor of such disorder. AlphaFold often draws an orange region as a long, loose ribbon, and this ribbon does not represent a real structure.
4:34 A real AlphaFold model
Fourth, with the bands in mind, we can read the scores of a real model from the AlphaFold Database. This model is the nucleocapsid protein of SARS-CoV-2, a protein of 419 amino acids that packages the viral RNA genome inside the virus particle. Its two folded domains appear in blue. The domain near the N-terminal end binds RNA, and the domain toward the C-terminal end holds two copies of the protein together as a dimer. Along the chain, the pLDDT of the two domains averages about 90 and 89. Three other regions, at the two ends of the chain and in the linker between the domains, average below 50, with means between 31 and 41. Experiments with single-molecule fluorescence show that all three regions are intrinsically disordered.
5:23 When the scores mislead
Fifth, the nucleocapsid example shows low scores that agree with experiments, but a score does not always have such a simple cause. Low pLDDT has two possible causes. The region may be intrinsically disordered, or it may have a real structure that AlphaFold lacks the information to predict. Much of that information comes from related sequences, which AlphaFold reads as a multiple sequence alignment, with the related sequences lined up position by position. The depth of the alignment at a position is the number of effective sequences there, where near-identical sequences count less than distinct sequences. The accuracy of AlphaFold2 drops substantially when the median depth falls below about 30 sequences. A pLDDT value on its own cannot separate these two causes, so you need further evidence, such as the depth of the alignment, a disorder predictor or an experimental structure of a related protein. A high pLDDT can also mislead, because some regions are disordered in the isolated protein but fold when they bind a partner or after phosphorylation, the addition of a phosphate group. AlphaFold2 tends to predict such regions in their folded state, with high scores.
6:37 Predicted aligned error
Sixth, pLDDT describes each residue only within its local neighborhood, so the arrangement of domains needs a separate score. Both domains of the nucleocapsid protein score high, yet the model does not show reliably where the two domains sit relative to each other. For this question, AlphaFold reports the predicted aligned error, or PAE, for each pair of residues. If you align the predicted and true structures on the first residue of the pair, the PAE is the expected error, in angstroms, in the position of the second residue. AlphaFold estimates this error during the prediction, without a true structure to compare against. Low values mean that AlphaFold is confident about where the two residues sit relative to each other. The PAE plot is a square grid with residue numbers along both axes, and darker squares mean lower error. In the nucleocapsid model, the squares within each domain are dark, at about 4 angstroms. The squares between the two domains are pale, at about 27 angstroms, close to the top of the scale. AlphaFold therefore has no reliable information on where the domains sit relative to each other, and experiments show that the two domains, joined by a disordered linker, interact only weakly. Before you measure a distance between two domains, or place a small molecule in the gap between them, check the PAE for that pair of regions.
8:01 Scores for protein complexes
Seventh, the scores so far describe a single protein chain, and many research questions concern molecules bound together. A complex forms when two or more chains bind each other, such as two proteins or a protein and DNA. AlphaFold 3, released in 2024, predicts complexes that contain proteins, nucleic acids, ions and small molecules. Two further scores build on the TM-score, which measures the similarity of two whole structures on a scale from 0 to 1. The predicted TM-score, or pTM, estimates the TM-score for the whole prediction, and a pTM above 0.5 means the overall predicted fold might be similar to the true structure. Where two chains touch, they form an interface, and the interface predicted TM-score, or ipTM, estimates how accurately AlphaFold 3 has placed the chains relative to each other. An ipTM above 0.8 indicates a confident prediction, and an ipTM below 0.6 indicates a likely failed prediction. The range from 0.6 to 0.8 is a gray zone, in which the score alone cannot show whether the prediction is correct.
9:13 Checking a model yourself
Finally, with all the scores in place, we can check a model ourselves and see how other software uses pLDDT. In an experimental structure, the B-factor of each atom describes how much the position of that atom varies around its average. Structure files keep the B-factor in a column of its own, and AlphaFold Database model files store pLDDT in that column instead. Molecular viewers such as ChimeraX can color a model by this column with the AlphaFold palette, so the four bands appear on your own screen. Suppose you open an AlphaFold model file and see a B-factor of 85 for a residue in a binding site. In an AlphaFold file, that value is the pLDDT, and a score of 85 falls in the high band, where the backbone is generally right but the side chain may be out of place.
10:03 Review and question
To review, pLDDT estimates how well AlphaFold has predicted the local structure around each residue, on a scale from 0 to 100. The AlphaFold Database colors the score in four bands, and scores below 50 usually mark disordered regions. The PAE describes the relative position of two residues, so it decides whether you can trust the arrangement of domains. For complexes, AlphaFold 3 adds pTM for the whole prediction and ipTM for the interface between chains. Here is a question to test what you learned. Two domains of a predicted structure are both dark blue, and the PAE between them is about 27 angstroms, so can you trust the distance between the two domains? Please write your answer in the comments below.