Do you really understand what the pLDDT score in protein structures means?

Researchers working in protein engineering and enzyme engineering deal with pLDDT almost every day. After running AlphaFold, we first look at the overall score distribution; when ESMFold outputs results, we glance at the confidence near the active site. But if someone asks, 'What's really the difference between pLDDT 70 and pLDDT 85?' not many people can explain it clearly. Some treat it as the gold standard for structure quality, some give up on the entire sequence if it’s below 70, and others directly compare pLDDTs from different tools—behind all these practices lies the same question: do you really understand what the pLDDT score in protein structures means? Why do some regions have pLDDT above 95 while others barely reach 40? Does a low pLDDT automatically mean the model prediction is wrong? When doing mutation design or enzyme modification, how should pLDDT actually be used? Today, we're going to clarify all these questions thoroughly.

1. What exactly is pLDDT?

The full name of pLDDT is predicted Local Distance Difference Test. It’s a per-residue confidence score introduced in AlphaFold2, used to measure how trustworthy the predicted position of each amino acid residue is in the structure.

Simply put, the core meaning of a protein structure's pLDDT score isn’t about asserting 'whether this structure is right or wrong,' but about 'how confident the model is in the predicted position of this residue.' Its value ranges from 0 to 100—the higher the score, the more reliable the model considers that residue's predicted position.

Here’s an important point: pLDDT is a local metric, not a global one. A protein chain may have half of its regions with pLDDT > 90 (very reliable), and the other half < 50 (highly uncertain). This often reflects the protein’s real biological characteristics—there’s a rigid structural core and flexible disordered regions. Understanding this is the starting point for using pLDDT correctly.

2. The real meaning of pLDDT score ranges

pLDDT confidence score intervals

Many people remember the "70-pass line," but what each range actually means and how to use it in real research isn’t always clear. It’s worth noting that the following range divisions are approximate references based on DeepMind’s official suggestions and domain practice experiences, not absolute judgment criteria.

pLDDT > 90: High-confidence region

This is usually the part the model is most certain about, often corresponding to the stable core regions of the protein—like hydrophobic cores, β-sheets, or tightly packed α-helices. The backbone predictions in these regions are generally quite reliable, and the side-chain conformations in the core areas are mostly of high reference value. For enzyme engineering, if the pLDDT near the active site is high, then molecular docking and mutation design based on this structure are relatively trustworthy.

70 ≤ pLDDT ≤ 90: Medium-confidence region

This typically covers most protein surface loops and flexible regions. It’s generally thought that the backbone direction is basically correct, but there may be deviations in the details—such as side-chain orientations and the exact conformation of loops. When analyzing protein interaction interfaces, if an interface falls in this range, it’s usually recommended to cross-check with other evidence.

50 ≤ pLDDT < 70: Low-confidence region

Predictions in this range can only provide a rough idea of the general direction. The backbone may have an approximate outline, but the model’s grasp of specific folding details is low. These regions are usually flexible loops or connecting segments on the protein surface, or partially intrinsically disordered regions (IDRs). If you’re studying a flexible linker, it’s common for its pLDDT to fall in this range—this is something easily overlooked when interpreting protein structure pLDDT scores: a low score doesn’t necessarily mean the model is inaccurate; it could also reflect the nature of the region itself.

pLDDT < 50: Very low confidence / potentially disordered

This score range typically indicates one of two possibilities: either the region naturally lacks a fixed 3D structure (intrinsic disorder), or the model can’t predict it accurately due to insufficient homologous sequences or complex folding. Many proteins’ N-terminal, C-terminal, signal peptides, or transport peptides are indeed naturally disordered and dynamic in solution. In this case, a low pLDDT is a valuable hint, but it’s recommended to use disorder prediction tools like IUPred or DisoPred and experimental data for confirmation. You can’t simply conclude that the region is disordered based solely on pLDDT < 50.

3. How is pLDDT calculated?

Understanding what the pLDDT score means for protein structures isn't just about memorizing thresholds—you also need to know the underlying logic.

When AlphaFold predicts 3D structures, it also estimates the local accuracy of each residue. Essentially, pLDDT is the model's direct prediction of the per-residue lDDT-Cα score (originally ranging from 0-1), multiplied by 100 to get a 0-100 score, which estimates how consistent the predicted local structure is with the experimental one.

Specifically, IDDT (Local Distance Difference Test) is a structural similarity metric proposed in 2013 that doesn’t rely on global structure superimposition. It was originally used to compare predicted and experimental structures. The original IDDT metric can be calculated based on different sets of atoms (like IDDT-all for all heavy atoms or IDDT-Cα for just Cα atoms). AlphaFold2 uses IDDT-Cα as one of its prediction targets. During training, the model learns to estimate its own prediction error and outputs per-residue predicted IDDT-Cα (i.e., pLDDT) as a confidence score.

One thing to note: pLDDT is a model’s "self-assessment," and it’s not foolproof. For some rare, new fold types not seen in the training set, the model can be "overconfident," giving high scores to predictions that aren’t actually very accurate. The opposite can also happen, where it’s "underconfident." So, pLDDT is an important reference, but it’s not an absolute gold standard.

4. Practical uses of pLDDT in protein and enzyme engineering

Applications of pLDDT

At the beginning, we mentioned that some people treat pLDDT as the gold standard, and some give up on an entire sequence as soon as they see it below 70—the root of these misunderstandings is precisely not knowing how to actually use pLDDT. After so much theory, let's get back to the question most relevant to experimentalists: how should we actually use pLDDT?

1. Quickly assess structural usability

When you get a predicted structure, first look at the overall pLDDT distribution. If the core regions (especially the active site) have pLDDT scores above 80, this structure can basically be used as a reliable basis for further calculations. If the pLDDT around the active site is only in the 60s, be cautious—it may require additional homology modeling or molecular dynamics simulations for further evaluation.

2. Identify naturally disordered regions

When doing enzyme immobilization or tag design, if you find that a certain terminal region has especially low pLDDT, it’s likely to be flexible. Adding tags or immobilization modifications there might have little impact on the overall protein structure. Conversely, if you want to do mutations in the core region, always prioritize positions with high pLDDT.

3. Guide mutation design priorities

When performing site-directed mutations, pLDDT can be a selection criterion. For surface residues, positions with high pLDDT are more structurally certain, making structure-based predictions of mutation effects more reliable. Low pLDDT regions are inherently flexible, so structural impacts of mutations are harder to predict accurately, and experimental verification is more uncertain.

4. Evaluate protein-protein interaction interfaces

When analyzing enzyme oligomerization or predicting protein complexes, changes in confidence at the interface can be somewhat informative. Regions with low pLDDT in monomer predictions that show increased confidence in a complex model may indicate that this region becomes stabilized upon partner binding. However, this shouldn’t be taken as independent proof of interaction and still needs to be combined with other analyses or experimental validation.

5. Common Misconceptions About pLDDT

The everyday mistakes mentioned earlier actually correspond to some long-standing misconceptions. In this section, we’ll clear them up.

Misconception 1: The higher the pLDDT, the more stable the protein

Not true. pLDDT measures prediction confidence, not the thermodynamic stability of the protein. An unstable mutant can still have a high pLDDT if its structure is well-defined. Conversely, a very stable protein may have low pLDDT in some naturally flexible regions.

Misconception 2: Regions with low pLDDT are useless

Completely wrong. Many important functional regions—like protein interaction interfaces, flexible enzyme lids, or post-translational modification sites—can fall in low-pLDDT flexible areas. The dynamics of these regions are part of their function. Low pLDDT shouldn’t be a reason to ignore them.

Misconception 3: pLDDT can be directly equated to experimental resolution

Not exactly. Studies show a general negative correlation between pLDDT and experimental structure resolution: regions with pLDDT>90 usually correspond to higher-resolution experimental structures, while 70-90 roughly aligns with medium resolution. But this is just a rough guide, not a precise conversion, as there’s a lot of variation across different proteins.

Misconception 4: pLDDTs from different tools can be directly compared

Not necessarily. Many structure prediction tools output pLDDT-like scores, but different models have different training methods and calibration strategies. An AlphaFold pLDDT of 70 might not mean exactly the same thing as an ESMFold pLDDT of 70. Keep this in mind when comparing across tools.

6. When pLDDT Meets AI: The Efficiency Revolution in Protein Engineering

Now that we understand what pLDDT scores mean for protein structures, let’s see how AI tools are pushing this forward. Here we need to mention MatwingsVenus™ (Xiaowu™)—an AI research assistant specifically designed for protein and enzyme engineering.

MatwingsVenus™

The core advantage of MatwingsVenus™ (XiaoWu™) lies in the fact that it doesn’t just help you predict structures and check pLDDT scores—it deeply integrates the pLDDT metric into the entire protein design process. From sequence input, structure prediction, and pLDDT confidence evaluation to active site identification, mutation design, and stability prediction, all the way through literature search, patent analysis, and experimental design, the MatwingsVenus™ (XiaoWu™) agent can handle everything in one place. It knows when to trust regions with high pLDDT scores and when to be cautious with low-score areas, automatically adjusting subsequent calculation strategies based on confidence. Even more impressively, it can proactively help you avoid the "minefield" of wet-lab experiments—for example, by automatically identifying regions near active pockets with lower pLDDT and suggesting you verify local conformational dynamics first using methods like NMR or HDX-MS, rather than blindly guessing during the computational stage. For researchers constantly dealing with structure prediction and mutation design, it’s like having a senior computational biologist right in your toolbox.

7. Between Certainty and Uncertainty: the Scientific Perspective pLDDT Teaches Us

The pLDDT metric may seem simple, but it actually hides a lot of underlying logic in protein structural biology. Truly understanding what protein structure pLDDT scores mean isn’t just about remembering a few thresholds; it’s about realizing that high-score regions tell you what is relatively certain, while low-score regions hint at what might be flexible or variable—and exactly where protein functions hide is often between this certainty and uncertainty.

Next time you open your structure prediction results, don’t rush to close them. Take a careful look at the pLDDT distribution curve. Those ups and downs in scores are actually telling you a story about how the protein folds, moves, and carries out its function. If you want to read this story faster and more accurately, consider letting MatwingsVenus™ (XiaoWu™) help you link pLDDT interpretation, mutation risk assessment, and experimental plans—it can finally let you spend the time worrying less about "is this score accurate" and more on designing the next breakthrough mutation.