LSMA

LONDON SCHOOL OF MULTIOMICS & AI

Free course

Free course

0 of 73 lessons

Lesson 8 of 12

Module 8 · AI for biology

AI structure prediction: what it changed, and its limits

AI structure prediction: what it changed, and its limits

AI structure prediction: what it changed, and its limits

Predicted structures are now available for almost every known protein. Learn to read the confidence scores first, download a real model and see which parts of it you can trust.

⏱ 16 min readModule 8 · AI for biologyFree
Lesson at a glance7 steps · about 16 min
  1. The folding problem
  2. The AlphaFold era
  3. Confidence scores: pLDDT and PAE
  4. Disordered regions
  5. Download a model and read its confidence
  6. Complexes
  7. What a prediction cannot tell you

The folding problem

A protein is made as a chain of amino acids, and in most cases that chain folds into a precise three-dimensional shape. The shape decides what the protein does. The folding problem is how to get from the sequence, which is cheap to read, to the shape, which is expensive to measure.

For decades, structures came only from experiments: X-ray crystallography, NMR and, more recently, cryo-electron microscopy. Each could take months or years. You met the result in File formats you will meet: FASTA, GenBank, GFF, VCF, PDB: the PDB file, one line per atom. The Protein Data Bank now holds more than 200,000 such structures, a small fraction of the proteins whose sequences we know.

Why should the sequence be enough? In the 1960s, experiments on the enzyme ribonuclease showed that an unfolded chain could fold back into its working shape on its own, so the information must be in the sequence. Reading it out is the hard part. A chain of a few hundred residues could in principle take an astronomical number of shapes, and the energy differences between the right fold and many wrong ones are small. In Protein structure: from sequence to fold you saw the levels of structure, from helices and strands to the whole fold; the folding problem asks a computer to reach the last level from the first.

Since 1994 the field has tested itself in CASP (Critical Assessment of protein Structure Prediction), a blind competition every two years: groups predict structures that are solved but not yet released, so nobody can see the answers.

CASP scores each prediction against the experimental structure, most often with GDT (global distance test), a score from 0 to 100 that counts how many residues sit close to their true positions after the two are superimposed. Blind testing matters because a method judged on structures it has already seen will look better than it is, the same leakage problem you met in A first sequence classifier, evaluated honestly.

The AlphaFold era

At CASP14 in 2020, the AlphaFold2 system predicted many targets with accuracy close to that of experiments. Three ideas, in plain terms, made the difference:

  • Evolution as evidence. The model reads a multiple sequence alignment (Module 4) of the protein and its relatives. When two positions change together across evolution, they are often touching in the folded protein.
  • Learning from known structures. A deep neural network was trained on the experimental structures in the PDB, so it learned what real proteins look like.
  • Refining a whole structure at once. The network repeatedly refines a full 3D model, and its own estimate of how reliable each part is.

The first idea is the easiest to picture. Suppose position 30 is a small residue that packs against a large one at position 85. If a mutation in some relative makes position 30 larger, the protein only stays folded if position 85 becomes smaller at the same time. Across thousands of related sequences, such paired changes leave a statistical trace, and that trace hints at which residues touch. The consequence is a common pitfall: a protein with few known relatives, such as one found only in a single group of organisms, gives a shallow alignment and usually a less confident model. The alignments you built in Module 4, Comparing sequences, are, in this sense, part of the input to every AlphaFold prediction.

In 2021 the AlphaFold Protein Structure Database (AlphaFold DB), run by EMBL-EBI, began releasing predictions; it now holds more than 200 million, covering most of UniProt. Other tools followed: RoseTTAFold, and ESMFold, which predicts structure from a protein language model like the one in Protein language models and embeddings without needing an alignment. Newer systems, including AlphaFold 3, also predict complexes of proteins with DNA, RNA, small molecules and ions. In 2024 the Nobel Prize in Chemistry recognised this work on structure prediction, together with computational protein design.

Cite AlphaFold 2 as Jumper et al. 2021, Nature, and the database as Varadi et al. 2022, Nucleic Acids Research. AlphaFold 3 is Abramson et al. 2024, Nature. When you use a model from AlphaFold DB, also record the UniProt accession, the file version (such as v6) and the download date, because models are updated. A predicted model is not an experimental structure, and should not be described as "the structure" of a protein. Reported accuracy comes from comparison with experimental structures, most of them of well-folded proteins, so confidence scores are least tested for unusual proteins and complexes.

Confidence scores: pLDDT and PAE

Every AlphaFold model comes with two confidence measures. Learn to read them before you look at the shape.

pLDDT (predicted local distance difference test) is a score from 0 to 100 for each residue. It estimates how well the local environment of that residue, its distances to nearby atoms, would match an experimental structure. AlphaFold DB uses four bands: above 90 is very high, 70 to 90 confident, 50 to 70 low, and below 50 very low. In the PDB-format file, pLDDT is stored in the column that normally holds the B-factor, the experimental measure of how much each atom moves. The column is reused, but the meaning is different.

This reuse catches people out in molecular viewers. Colouring an experimental structure "by B-factor" highlights the most mobile atoms with high values. Colouring an AlphaFold model the same way highlights the most confident residues with high values, the opposite reading. Always check whether a file is a prediction or an experiment before you interpret that column, and say which in any figure legend.

PAE (predicted aligned error) answers a different question: if you line up the model and the true structure on residue y, how far off, in ångströms (Å, one ten-millionth of a millimetre), would residue x be? It is shown as a square map with one row and one column per residue. Low PAE between two domains means their relative position is reliable. High PAE means each domain may be right on its own while the angle between them is unknown. The figure later in the lesson shows how to read one.

pLDDTBandHow to treat it
> 90Very highBackbone and most side chains likely right; fine for looking at a binding pocket.
70–90ConfidentBackbone usually right; side chains less certain.
50–70LowTreat the shape with caution; overall path may be wrong.
< 50Very lowDo not interpret the shape; often a sign of disorder.

How to read the table. Read the left column for the score of the residues you care about, then the right column for how much weight the shape can bear. And the score is local. A protein can have every residue above 90 while its two domains point the wrong way, which is what PAE is for.

Disordered regions

Many proteins contain intrinsically disordered regions: stretches that never settle into one fixed shape on their own. They are common in regulatory proteins. In AlphaFold models they come out as long, loose ribbons with very low pLDDT, which is one of the better computational signals that a region is disordered. Some disordered stretches fold only when they bind a partner.

Disorder is not a defect. A flexible region can reach out to several partners, can be modified quickly by enzymes that add phosphate groups, and can be unfolded and destroyed easily when the cell no longer needs it. The N-terminal region of p53 is an example: a short part of it forms a helix when it binds the partner protein MDM2, which controls how much p53 the cell keeps.

The pitfall runs the other way too. Low pLDDT does not prove disorder. A region can score low because it is ordered only inside a complex, or because too few relatives exist to give the model evidence. Before you write "disordered", look for support from dedicated disorder predictors or from experimental data, such as the curated records in the DisProt database.

Download a model and read its confidence

This lesson needs Biopython (!pip install biopython in Colab). The code downloads the AlphaFold DB model of human p53, UniProt P04637, which you met in Module 2. The file name follows the pattern AF-<UniProt accession>-F1-model_v6.pdb, where F1 means the first fragment (long proteins are split) and v6 the database version. The version number changes over time; the protein's page on the AlphaFold DB site always links to the current file.

Python
import urllib.request
from Bio.PDB import PDBParser

url = "https://alphafold.ebi.ac.uk/files/AF-P04637-F1-model_v6.pdb"
urllib.request.urlretrieve(url, "AF-P04637.pdb")

structure = PDBParser(QUIET=True).get_structure("p53", "AF-P04637.pdb")
plddt = {}                                     # residue number -> pLDDT
for residue in structure[0]["A"]:
    ca = residue["CA"]                         # the alpha-carbon atom
    plddt[residue.id[1]] = ca.get_bfactor()    # pLDDT sits in the B-factor column

print("Residues:", len(plddt))
print("Mean pLDDT:", round(sum(plddt.values()) / len(plddt), 1))
print("Residues 1-5:", [round(plddt[i], 1) for i in range(1, 6)])
print("Residues 240-244:", [round(plddt[i], 1) for i in range(240, 245)])

What this does

  • urllib.request.urlretrieve downloads the file from the URL and saves it as AF-P04637.pdb in the notebook's folder.
  • PDBParser(QUIET=True) reads PDB files; QUIET=True hides warnings about minor format quirks. get_structure returns a nested object: structure, then model, then chain, then residue, then atom.
  • structure[0]["A"] takes the first model and chain A. The loop visits each residue, picks its alpha-carbon ("CA", the central backbone atom of every amino acid) and reads its B-factor, which here holds the pLDDT. Every atom of a residue carries the same value, so one atom is enough.
  • residue.id[1] is the residue number. The dictionary plddt maps each number to its score.
  • The last lines print the count, the mean, and the scores of two short stretches: the first five residues and residues 240–244.
Residues: 393 Mean pLDDT: 75.1 Residues 1-5: [42.7, 43.6, 37.8, 45.9, 39.3] Residues 240-244: [98.2, 97.8, 97.4, 96.6, 93.8]

Reading the output

The model covers all 393 residues of p53. The mean pLDDT is 75.1, the same figure AlphaFold DB shows as the model's overall confidence. A mean hides a lot, though: the first five residues score around 40, very low, while residues 240–244, in the middle of the DNA-binding domain, score 94 to 98. The next block sorts every residue into its band and finds the confident stretches.

Python
bands = {"very high (>90)": 0, "confident (70-90)": 0,
         "low (50-70)": 0, "very low (<50)": 0}
for score in plddt.values():
    if score > 90:   bands["very high (>90)"] += 1
    elif score > 70: bands["confident (70-90)"] += 1
    elif score > 50: bands["low (50-70)"] += 1
    else:            bands["very low (<50)"] += 1
for name, n in bands.items():
    print(f"{name:18} {n:4} residues  {100 * n / len(plddt):5.1f}%")

# stretches of at least 10 residues in a row with pLDDT >= 70
stretches, run = [], []
for i in sorted(plddt) + [None]:            # None marks the end of the chain
    if i is not None and plddt[i] >= 70:
        run.append(i)
    else:
        if len(run) >= 10:
            stretches.append((run[0], run[-1]))
        run = []
print("Confident stretches:", stretches)

What this does

  • bands is a dictionary of four counters. The if / elif chain puts each score in the first band it fits, from the top down.
  • The f-string {name:18} pads each name to 18 characters, so the columns line up; {…:5.1f} prints the percentage with one decimal place.
  • The second loop walks along the chain, collecting residue numbers into run while the score stays at 70 or above. When the score drops, a run of at least 10 residues is saved as a (start, end) pair, and the run is emptied. The None added at the end forces the last run to be checked.
very high (>90) 206 residues 52.4% confident (70-90) 27 residues 6.9% low (50-70) 43 residues 10.9% very low (<50) 117 residues 29.8% Confident stretches: [(94, 182), (186, 292), (324, 355)]

Reading the output

Just over half the residues (52.4%) are very high confidence, and almost a third (29.8%) are very low. Three confident stretches stand out. Residues 94–182 and 186–292 are the DNA-binding domain, split only by a three-residue dip to about 69, just under the cut-off. Residues 324–355 are the tetramerisation domain, the part that lets four p53 chains join together. Everything else, including the N-terminal region that switches on target genes and the C-terminal tail, is low or very low: the regions known to be disordered.

Notice how much the result depends on two choices made in the code: the cut-off of 70 and the minimum run of 10 residues. With these settings the DNA-binding domain splits into two stretches, only because three residues dip to about 69. Lower the cut-off by one point and it becomes one stretch; raise the minimum run to 40 and the tetramerisation domain, 32 residues long, disappears. Thresholds like these are useful for a first summary, but always look at the scores around each boundary before you report domain limits, and state the rule you used.

AI structure prediction: what it changed, and its limits · LSMA free course A bar runs from residue 1 to 393 of p53, coloured by pLDDT band from the downloaded model. Residues 94 to 292, the DNA-binding domain, and 324 to 355, the tetramerisation domain, are mostly very high confidence. The N-terminal region, the linker from about 293 to 323 and the C-terminal tail after 356 are mostly low or very low confidence, which marks them as likely disordered. A short stretch near residues 20 to 25 is confident. AF-P04637-F1 · pLDDT PER RESIDUE 194292324356393 DNA-binding domain tetramer N-terminal region folded, pLDDT mostly above 90 domain mostly low or very low very high (>90) confident (70–90) low (50–70) very low (<50) Bands drawn from the pLDDT values read in the code below. Mean pLDDT 75.1 over 393 residues.
Confidence is not spread evenly: two folded domains stand out against long, low-confidence stretches. Read low confidence here as "probably disordered", not as "badly predicted".

How to read the figure. The bar is the protein from residue 1 on the left to 393 on the right, drawn to scale. Each colour is a pLDDT band from the legend. Dark navy means trust the shape; vermilion means do not. Brackets under the bar mark the regions named in the text. The short confident patch near residues 20–25 is a helix the N-terminal region forms when it binds a partner protein.

Complexes

Most proteins work with partners. AlphaFold DB holds single chains, so the p53 model shows the tetramerisation domain as one lonely helix-and-strand unit, not as part of the four-chain assembly it really forms. Tools for complexes, such as AlphaFold-Multimer and AlphaFold 3, take several chains as input and give an extra score for the interface between them. A confident-looking complex of two proteins that never meet in the cell is still possible.

So a predicted complex is best treated as a hypothesis about an interaction you already have reason to believe in, for example one supported by experiments or listed with high confidence in an interaction database of the kind used in Biological networks and systems biology. Predicting every possible pair and keeping the confident ones finds some real interactions, but among thousands of pairs some confident-looking false ones will slip through, for the same reason that many tests produce false positives.

AI structure prediction: what it changed, and its limits · LSMA free course A schematic square map with residues on both axes. Two dark squares on the diagonal show low error within domain A and within domain B. The off-diagonal blocks between the domains are pale, meaning high error: each domain is reliable on its own, but their relative position is not. A note on the right explains the three readings. PREDICTED ALIGNED ERROR · SCHEMATIC domain Adomain B aligned residue (y) scored residue (x) low errorlow error high errorhigh error Dark block on the diagonal the domain's own fold is reliable Pale block off the diagonal where A sits relative to B is uncertain Whole map dark domain arrangement is reliable too
PAE is about relative positions. Each domain here is trustworthy, but the hinge between them is not, so do not read anything into the angle between A and B. Schematic, not from a real protein.

How to read the figure. Find the residue you are scoring on the left axis and the residue you align on along the bottom. The colour where they cross is the expected error: dark is small, pale is large. On AlphaFold DB, each entry shows the real PAE map for that protein next to the 3D view.

Go deeper: reading PAE and interface scores for complexes

Two details trip people up. First, a PAE map is not symmetric. The error for residue x when aligned on y can differ from the error for y when aligned on x, so the two halves of the map, above and below the diagonal, need not match. For a complex, the off-diagonal blocks between two chains are the ones that tell you whether their placement is reliable.

Second, complex predictors give whole-model scores. pTM estimates how well the overall fold matches, from 0 to 1. ipTM does the same for the interfaces between chains only. Guidance for AlphaFold 3 suggests that ipTM above about 0.8 indicates a confident prediction, below about 0.6 a likely failure, and values in between a grey zone. Large disordered regions can drag these scores down even when the interface itself is good, so read them together with the PAE map.

What a prediction cannot tell you

  • Movement. A model is one static shape. Proteins flex, open and close; enzymes and channels often work by changing shape. A single prediction does not show these states or how often each occurs.
  • Ligands, ions and modifications. AlphaFold DB models contain only the protein chain. The p53 DNA-binding domain holds a zinc ion that is essential for its fold, and it is not in the file. Bound drugs, sugars and phosphate groups are missing too.
  • Effects of mutations. The model was not trained to tell how one amino-acid change alters stability or function. Predicting a mutant usually returns almost the same structure, even for changes that destroy the protein in the cell. Use dedicated variant-effect tools and experiments for that question.

None of this makes predictions less useful; it says what they are useful for. In practice, researchers use models to decide which part of a protein to express in the laboratory, to choose which residues to mutate, to locate a likely binding site, and to help solve experimental structures, where a good model can serve as a starting point for interpreting X-ray or cryo-EM data. In each case the model guides an experiment rather than replacing it.

Worked example · is this region worth studying in the model?

region A: residues 240–244 · pLDDT 93.8–98.2
region B: residues 1–5 · pLDDT 37.8–45.9
p53 overall: 206 of 393 residues very high (52.4%), 117 very low (29.8%)

A colleague wants to use the p53 model to see where residues 240–244 sit, and also to describe the shape of the first five residues.

Step by step. Look up each region in the output. Region A scores from 93.8 to 98.2, all above 90, so it is in the very high band and lies inside the confident stretch 186–292. Its position in the domain can be trusted. Region B scores from 37.8 to 45.9, all below 50, in the very low band, at the start of the long N-terminal region. Check the overall split too: 206 ÷ 393 = 0.524, so about half the protein is very high, and 117 ÷ 393 = 0.298, so nearly a third is very low.

What the answer means. Region A can be studied in the model, for example to see which residues face the DNA. Region B should not be described as having any particular shape; the right statement is "predicted to be disordered". Neither answer says anything about the zinc ion, about the tetramer, or about what a mutation at residue 240 would do.

Key point

Read the confidence before the shape. pLDDT tells you which residues to trust, PAE tells you whether domains are placed correctly, and very low confidence often means disorder. A prediction is one static chain, without its ligands, partners or movements.

A second example: a predicted model of a membrane channel looks beautiful, with all residues above 90. It is still one shape, with no membrane, no ions and no information about whether the pore is open or closed. Whether the channel conducts in that state needs experiments or simulations, not a second look at the model.

Further reading · Module 8AlphaFold database FAQ: pLDDT, PAE and file versions. · EMBL-EBI AlphaFold training course. · PDB-101 guide to understanding PDB data. · All reading for this course

Check your understanding

1. A residue in an AlphaFold model has a pLDDT of 38. What is the most sensible reading?

2. Every residue of a two-domain protein scores above 90, but the PAE between the two domains is high. What can you say?

3. A colleague wants to use an AlphaFold model to decide whether a missense variant destabilises a protein. Why is that a poor use of the model?

Up next

Lesson 9

AI for multi-omics and single-cell data

Ask BioTutor

Create a free website with Framer, the website builder loved by startups, designers and agencies.