Skip to content
PaperFren

Protein folding · Deep learning · Generative models

Protein structure predictors get the answer without learning how proteins fold

Evidence: PreliminaryOne study or a small sample; not yet replicated. What the labels mean

Study published Jan 1, 2022. PaperFren added this explanation Sep 26, 2026.

Save this development or follow its topic to track what changes.

Short answer

Deep structure predictors reach accurate final structures without implicitly capturing folding physics, and generative models learn their training data's errors along with its patterns.

What happened

Researchers ran structure predictors on 170 proteins with measured folding kinetics and treated their search trajectories as folding pathways. All methods beat chance at classifying folding type, but none beat chain length alone; folding-rate correlations were mostly non-significant or wrong-signed, and predicted intermediates were mostly no better than random, with over 80% containing clashes for some methods. Separately, a diffusion model trained on simulations of disordered peptides reproduced its reference ensembles quickly but underestimated experimental size (RMSE 0.70 nm), matching the simulations' own bias.

Why it matters

Accurate final structures invite the reading that a network has learned the underlying physics. These papers separate the two: predicting an endpoint well is not the same as modelling the process, and a model trained on simulations learns the simulations, not nature.

Evidence

Study type
Computational comparison against experimental folding kinetics; generative-model benchmark
Sample
170 proteins with folding kinetics (79 for rate analysis, 9 for the AlphaFold 2 comparison); over a thousand simulated peptides and 10 experimental test peptides
Journal
Bioinformatics · peer reviewed
Replication
Single study for the folding-pathway question; not yet repeated with newer predictors
Limitations
Search trajectories are an indirect probe of learned physics; the result says nothing against final-structure accuracy. The disordered-protein model fails beyond 60 residues and its bias fix breaks chains.

What this connects to

Sources

The 2 studies this explanation is built from, by the role each plays. Every source links to PaperFren’s explanation of it and to the original paper.

Primary study

  • Do protein structure predictors understand how proteins fold?

    Outeiral C, Nissley DA, Deane CM · 2022 · Bioinformatics (Oxford, England) · 69 citations

    The step-by-step paths that modern protein structure predictors take to their answers do not match how real proteins fold, and a protein's length alone predicts folding behaviour better.

    What it does not show

    Structure predictors were never designed to simulate folding, so their search paths are an indirect probe; the result says they have not implicitly learned folding physics, not that they are poor at predicting final structures. AlphaFold 2 was tested with only one trajectory per protein, so its results could not be tested for significance. The experimental ground truth is itself fuzzy: whether a protein is 'two-state' depends on conditions like temperature and denaturant, and some classifications are disputed in the literature.

    PaperFren explanationStudy with cards and a quizOriginal paper (DOI)cc by

Supporting evidence

  • Can a diffusion model generate shapes of floppy proteins?

    Janson G, Feig M · 2024 · PLoS computational biology · 54 citations

    A diffusion model trained on simulations generated realistic shape ensembles for disordered peptides it had never seen, far faster than simulation, but it inherited the simulations' bias toward overly compact shapes.

    What it does not show

    The model learns simulations, not reality, so it inherits their errors such as overly compact disordered chains. It fails for sequences much longer than the 60-residue training maximum and for highly helical peptides rare in training. Larger networks gave no statistically significant gains, and the biased-sampling fix is unstable. The supplied text begins partway into the results, so the introduction and some method details are missing here.

    PaperFren explanationStudy with cards and a quizOriginal paper (DOI)cc by

Before

A network that predicts protein structures as well as AlphaFold must have learned something like the physics of folding.

Now

Their search paths do not match experimental folding behaviour. The caveat is important: these tools were never designed to simulate folding, AlphaFold 2 was tested with one trajectory per protein, and experimental ground truth on intermediates is itself fuzzy.