Case Notes

Helped: filling in the nuclear pore 2022

Documented The nuclear pore complex is the gate in the envelope around a cell's nucleus. Everything that moves between the nucleus and the rest of the cell, from RNA on its way out to proteins on their way in, passes through it. It is one of the largest protein assemblies in the cell. Its proteins, the nucleoporins, fall into two groups: scaffold proteins with folded domains, which build a ring-shaped wall, and disordered ones, which line the wall and reach into the central channel. ^[1]

Documented For years, electron microscopy had shown the overall shape of the pore at moderate detail, and X-ray crystallography had solved many single pieces. But many nucleoporins, and many of the places where they touch each other, had no experimental structure at all. The map had a shape; much of what filled it was missing. ^[1][2]

Documented Why was it so hard? The whole pore is far too big and too flexible to crystallise, it sits in a double membrane, and its shape changes: the abstract of the 2022 paper calls its architecture highly dynamic. So it has to be studied in pieces and put back together. ^[1]

Documented In 2022 a team led from the European Molecular Biology Laboratory and the Max Planck Institute of Biophysics used AI-based structure prediction to model the human nucleoporins and many of their sub-complexes, including pieces and contacts no experiment had shown. They benchmarked the models against X-ray and cryo-EM structures, some not yet published, and found them unusually accurate. ^[1]

Documented They fitted the models into new cryo-electron tomography maps of the pore in two states, narrowed and widened, taken inside cells. The result was a 70-megadalton model, in atomic detail, covering more than 90% of the human pore's scaffold. It was complete enough to run computer simulations of the scaffold sitting in a membrane. ^[1]

Documented The simulations gave a finding of their own: without tension in the membrane, the scaffold stops the pore in the double membrane from closing down to a small diameter. A question about the living pore could be asked only because the model was nearly complete. ^[1]

Documented A second paper in the same issue of Science combined cryo-EM and AlphaFold models to build the cytoplasmic ring of the pore, the part facing the rest of the cell. ^[3]

Documented What the model did not show matters too. The disordered nucleoporins that fill the central channel have no single shape, and a scaffold model does not give them one. Writing alongside the papers, Thomas Schwartz called the work solving the nuclear pore puzzle; the puzzle solved was the wall, not the gate's floppy lining. ^[1][2]

Documented What made the predictions useful was not that they were trusted. They were tested. Each predicted piece had to fit an experimental map that the prediction had never seen, and the pieces had to fit together. The prediction filled gaps; the experiment showed whether each fill was right. ^[1][3]

Reconstructed We have no record of the day-to-day work, so this is how such a fitting job goes. A predicted piece is placed into a blurry map like a part into a mould; a computer scores how well it fits; pieces that fit badly are set aside or remodelled, and a piece with low confidence is trusted less than one with high.

The tool that would have flagged it. The habit that made this work was checking each predicted piece against independent data, and checking how sure the model was about how the pieces sit together. Practise the second with HingeHunt; compare a model with an experiment in FoldCompare. HingeHunt · FoldCompare

Misled: proteins with two folds 2022

Documented Most proteins have one folded shape that moves a little. A few, called fold-switching or metamorphic proteins, have two quite different stable folds. The bacterial protein RfaH is a well-known case. Part of it, its C-terminal domain, is a pair of α-helices while it packs against the rest of the protein. When it is released, the same stretch of chain refolds into a β-barrel, and in that form it does a different job: it helps link the reading of a gene to the making of protein. ^[4]

Documented KaiB, a protein of the daily clock in cyanobacteria, is another. Most of the time it sits in one fold. Rarely, it switches to a second fold, and only in that form can it bind the clock protein KaiC. The slowness of the switch helps set the timing of the clock. ^[5]

Documented Fold switchers are thought to be uncommon, but they may be under-counted. A 2018 analysis of the Protein Data Bank estimated that somewhere between about 0.5% and 4% of proteins there could switch folds. ^[6]

Documented In 2022 Devlina Chakravarty and Lauren Porter at the US National Institutes of Health tested AlphaFold2 on 98 fold-switching proteins. For each, they compared five predicted models with the two experimental structures. In 94% of cases the predictions matched one of the two experimental folds and not the other. ^[7]

Documented The confidence did not warn of the problem. AlphaFold2's confidence was moderate to high for 74% of the fold-switching residues. That is different from disordered regions, where confidence is usually low, so a reader looking only at the colours would not have been alerted. ^[7]

Documented The authors suggest that AlphaFold2 does well on what is apparent from structures already solved, and poorly where a protein's behaviour is not, and that a protein is better thought of as an ensemble of shapes than as a single one. ^[7]

Disputed Can AlphaFold be coaxed into the second fold? Some groups report that running it in special ways, for example on clusters of related sequences, recovers alternative folds for some proteins. A 2025 follow-up from the same lab argues that many such successes rest on structures already present in AlphaFold2's training set, and that true prediction of fold switching is still weak. ^[8][9]

Reconstructed We have no record of any one person misled by such a model, so this is how such a misreading would go. A student opens the model of a fold switcher, sees a confident, well-packed structure, and writes that this is the shape of the protein. The model is not wrong about that fold. It is silent about the other one.

The tool that would have flagged it. No colour on the model gives this away; the warning is the habit of asking whether a protein has more than one known shape. NotMarbles practises it, and FoldCompare lets you lay a model beside each experimental structure in turn. NotMarbles · FoldCompare

Misread: a confident model of a part that is floppy on its own 2023

Documented A common rule of thumb says that where a protein is disordered, floppy and without one fixed shape, AlphaFold's confidence is low. It is often true, and it is why low confidence is so useful. ^[10]

Documented But in 2023 Reid Alderson, Julie Forman-Kay and colleagues showed that AlphaFold2 gives confident structures to nearly 15% of human disordered regions. Comparing with experiments on regions known to fold only in certain conditions, such as when bound to a partner or after a chemical tag is added, they found that AlphaFold2 often predicts the folded state. ^[11]

Documented So for these regions the model shows a shape the protein takes under some conditions, not the shape it has alone in water. The authors call this conditional folding, and found that such regions are enriched in disease-linked mutations. ^[11]

Documented The confident scores were useful as a signal: on the regions known to fold in some condition, high confidence picked them out with a precision of up to 88% at a 10% false-positive rate. The same study found conditionally folding regions far more common in bacteria and archaea than in eukaryotes, and concluded that most eukaryotic disordered regions do their jobs without taking up a fixed shape. ^[11]

Documented Alpha-synuclein, a small protein of nerve cells, is a clear example. Alone in solution it is largely unfolded. Bound to a lipid surface, most of its first hundred residues form helix: two helices on small detergent micelles, and one long, extended helix on larger membranes. ^[12][13][14]

Documented Alpha-synuclein is also the main protein in Lewy bodies, the clumps found in the brain cells of people with Parkinson's disease, which is one reason it is so widely studied, and so often shown. ^[15]

Documented In the AlphaFold DB model, AlphaFold is confident about a long helix over residues 1 to 95, and is much less sure about the tail from 96 to 140. A long helix is what the protein forms on membranes, not alone in solution. ^[16][14]

Documented A reader who takes the model as the protein's only shape would say alpha-synuclein is a stable helix. That is not what experiments on the free protein show. The model is a confident picture of one condition. ^[11][12]

Disputed What is the protein's 'real' shape? One view is that a disordered protein has no single shape and the model is a picture of a bound state. Another is that the bound state is the one that does the work, so the model shows what matters. Both views fit the data; they differ in what they call the protein's shape. ^[11][12]

Reconstructed We have no record of a particular misreading, so this is how one would go. Someone looking through a database for a protein to use as an example of a disordered protein skips alpha-synuclein because its model is so confident, and picks a less typical one instead.

The tool that would have flagged it. Again no colour flags it: the model is confident. The questions that catch it are 'what kind of picture is this?' and 'in which conditions?'. HonestReport asks the first; DisorderDuty practises the opposite case, where low confidence is the real answer. HonestReport · DisorderDuty