Understanding AlphaFold's pLDDT, PAE, and pTM

September 10, 2026 machine learning, structural biology, alphafold share

Proteins are not inherently good or bad. Fold correctly and they hold your muscles together, keep your skin intact, and give your bones their strength. Fold wrong, because of a single DNA mutation, and you get Cystic Fibrosis, Huntington's disease, or the rigid haemoglobin rods that cause sickle cell anaemia. The shape is the whole story, and until recently, finding it experimentally, by X-ray crystallography, NMR, or cryo-EM, could take a PhD student's entire career and run past $100,000 per structure. Even shifting the search to a computer ran into Levinthal's paradox: a chain flexible enough to have on the order of 10^43 possible shapes is too large a space to search one configuration at a time, even given more time than the universe has existed.

That was the situation before AlphaFold: an AI system from Google DeepMind that predicts a protein's 3D structure from its amino acid sequence alone, solving a fifty-year-old problem in biology and earning its creators a share of the 2024 Nobel Prize in Chemistry.

What I actually want to write about here isn't the model. It's the report card it hands back with every structure.

This is what AlphaFold Server gives you back: a structure colored by confidence, a pTM score, an ipTM score, and a green-toned grid next to it. pTM comes back 0.18. ipTM is a dash, not a number. The grid is pale almost everywhere except a thin dark line down the middle.

Read literally, none of that tells you whether the structure is right. It tells you what the network thinks of its own answer, and those are two different questions. Telling them apart means knowing what each number is actually testing.

What pLDDT actually is

lDDT, the Local Distance Difference Test, predates AlphaFold. Mariani and colleagues introduced it in 2013 as a way to score a predicted structure without first aligning it to the true one: for each atom, look at everything within 15 Å in the real structure, then check whether the predicted model keeps those same neighbors at roughly the same distances, within four tolerance bands (0.5 Å, 1 Å, 2 Å, 4 Å). Average the pass rate across those bands and you get a 0 to 1 score per atom for whether its immediate neighborhood came out right. Because it never needs a global alignment step, a wrong hinge angle between two otherwise-correct domains doesn't drag down the score for atoms sitting inside either one.

AlphaFold's version, pLDDT, predicted Local Distance Difference Test, is the network estimating what its own lDDT score would be if you could check it against a real structure. It learned to make that estimate from thousands of real structures during training, where the true answer was available. At inference time, on a new sequence, there is no true structure to check against. pLDDT is a self-assessment, not a measurement.

GFP's one real dip

Green fluorescent protein, avGFP, the original variant cloned from the jellyfish Aequorea victoria (UniProt P42212), is a clean place to see this, because most of it is easy: eleven beta strands wrapped into a barrel around a single central helix, folded the same way in every jellyfish that makes it. Running my own sequence through AlphaFold Server gives a pLDDT that sits in the high 90s for nearly the entire chain of 238 residuesOne amino acid unit in a protein's chain, numbered from one end to the other., averaging 92.8.

There's one real dip, and it lands exactly where you'd want it to. Residues 65 through 67, serine, tyrosine, glycine, drop from the high 90s down to 78.2, 70.7, and 74.9 before climbing straight back to the high 80s and 90s on either side. That tripeptide is GFP's chromophoreThe part of a molecule that absorbs light, here the group responsible for GFP's fluorescence., the three residues that fold in on themselves and undergo an internal chemical reaction to become the light-emitting group, a rearrangement unlike anything else in the chain. The network isn't confused about the fold there. It's telling you, correctly, that this one spot behaves differently from the rest of the barrel.

What PAE actually is

pLDDT asks a local question: is this one residue's immediate neighborhood right? Predicted Aligned Error asks a relational one: if you anchor residue i exactly in place and align the structure around it, how far off, in Ångströms, would residue j land? That's why PAE isn't a single number per residue but a full i by j matrix. Two domains can each be internally solid, each with excellent pLDDT throughout, while the PAE between them stays high, because the model has no idea how the two domains sit relative to each other. pLDDT alone would miss that entirely.

GFP's own PAE matrix is close to uniformly dark: a mean error of 3.6 Å across the whole grid, with 88% of it under 5 Å. That's what a single rigid domain looks like in this kind of plot, confident not just about each residue's own shape but about how every residue relates to every other one, all at once.

Beta-casein, built without a fixed shape

Beta-casein is a milk protein, and it isn't malfunctioning by lacking a fixed shape. It's built that way. Casein micelles hold together through loose, shifting contacts between disordered chains rather than any one protein folding into a rigid form. An intrinsically disordered protein like this one has no single correct structure for AlphaFold to find, because there isn't one to find.

Run it through AlphaFold Server and the numbers show exactly that. Confidence holds for the first fifteen residues, comfortably above 90 pLDDT, dips briefly at 16 and 17, then declines for good: 83% of the chain sits below 70, and a third sits below 50. pTM comes back at 0.18, against GFP's 0.91. The PAE matrix mirrors it: a mean error of 27.5 Å, with only 4% of the grid under 5 Å. Where GFP's plot is a dark square, beta-casein's is pale almost everywhere, with barely a shadow along the diagonal.

Rotate either one and the pattern holds from every angle: GFP's barrel keeps its shape, solid blue except for its two floppy termini, while beta-casein's chain never settles into anything you'd call a fold.

Under the hood: what the numbers are made of

pTM and ipTM aren't separate measurements. They're PAE, reduced to a single number. The predicted TM-score runs each entry of the PAE matrix through the same distance-normalizing function the original TM-score uses, then averages the result. A tight, low PAE matrix collapses into a high pTM; a diffuse one collapses into a low one, which is exactly why GFP lands at 0.91 and beta-casein lands at 0.18.

ipTM is the same calculation restricted to residue pairs that cross between chains, i in one chain, j in another. GFP and beta-casein were both run as single chains, so there's no inter-chain pair to average over, and ipTM comes back undefined, a dash on the Server's own page, not a zero and not a missing value. It only exists for interfaces.

AlphaFold Server also reports fraction_disordered, a third check that doesn't ask the network what it thinks of its own confidence at all. It's computed from the geometry of the predicted structure itself, using the kind of classical, pre-AlphaFold criteria structural biologists have long used to call a region disordered. And it feeds into the ranking formula in a way that's easy to miss:

ranking_score = 0.8 * iptm + 0.2 * ptm + 0.5 * fraction_disordered - 100 * has_clash

The fraction_disordered term is positive. A model that correctly identifies a region as disordered gets rewarded for it, not punished for failing to force a shape onto something that doesn't have one. That's a deliberate choice: treating every low-confidence region as a failure would push a model toward hallucinating structure into places like beta-casein's tail just to keep its score up.

Confident, not correct

GFP's dip at its chromophore and beta-casein's disorder are both real signals, and the network reported them accurately. But that's the catch: pLDDT, PAE, pTM, and ipTM are all the model grading its own prediction, not a comparison against a structure verified in a lab. They're worth trusting as an honest self-assessment. They aren't ground truth, and treating them as one is exactly how a confident-looking structure turns out to be wrong. The model and confidence code, including confidences.py, is open on GitHub.