Concept
Machine learning for protein structure prediction
6 studies1 discoveryEvidence last moved Sep 27, 2026
Protein structure prediction uses sequence information, especially coevolution between residue positions, to infer contacts and 3D shape, and then estimates how good predicted models are. The evidence here is from computational studies spanning SVM contact prediction, coevolution models, neural-network meta-predictors, deep model-quality assessment, and a test of whether modern predictors learned folding physics.
Modern predictors are extremely accurate, but it helps to know what signal they rely on and what they do not capture. These papers trace the path from contacts to models and show that predicting a final structure is not the same as understanding folding.
Studies
6
Findings
5
6 supporting · 0 challenging · 0 qualifying citations
Open tensions
1
Latest change
Concept page published
Machine learning for protein structure prediction
Currently
What we know
- Before coevolution methods, most long-range contact predictions were wrong.
- The right statistical model extracts structural signal from correlated mutations.
- Learned combination beats naive averaging of predictors.
- Estimating model quality is its own learning problem.
- Accurate end-point prediction does not mean the model learned folding physics.
Largest unresolved question
More precise contacts did not always give better 3D models: MetaPSICOV's more precise stage two produced slightly worse models than stage one because its extra contacts were redundant neighbours in beta sheets.
Common misconceptions
Because AlphaFold predicts structures well, it simulates how proteins fold.
Its search trajectories did not reproduce folding types, rates or intermediates better than chain length or chance; the result concerns folding physics, not final-structure accuracy.
Principal-component-like (large-eigenvalue) correlation patterns hold the structural signal.
Contact accuracy came almost entirely from small-eigenvalue repulsive patterns; using only PCA-like patterns sharply reduced it.
Related
Claim ledger
What the evidence shows
Drawn from 6 studies in this library. Mix labels say which citation roles are present; they are not a strength score. Supports means evidence for a finding; Challenges means evidence against a stated position; Qualifies marks scope.
Before coevolution methods, most long-range contact predictions were wrong.
Early machine-learning contact prediction was weak: SVMcon reached 27.1% accuracy at the break-even point, about nine times random, and in CASP7 ranked second among eight predictors.
The right statistical model extracts structural signal from correlated mutations.
Coevolution analysis finds contacts from sequence alignments; the low-eigenvalue ('repulsive') patterns in a Hopfield-Potts model carried nearly all contact accuracy, reaching 96% of top contacts in one family and 70–80% with only 10–30 sequences where DCA fell to about 30%.
Learned combination beats naive averaging of predictors.
Combining several coevolution methods with a neural network (MetaPSICOV) raised top long-range contact precision 38% over the best single method, while simply averaging them did no better; the network relied on non-coevolution features when alignments were shallow.
Estimating model quality is its own learning problem.
Deep models for model-quality assessment improved ranking modestly: DeepQA tied the leading single-model method on CASP11 stage one and led on stage two, and a later equivariant GNN beat AlphaFold2's own confidence score (0.90 vs 0.84 correlation) on AlphaFold2 models.
- Can deep learning judge how good a predicted protein structure is?
- Can an equivariant graph network judge protein models better?
Study Role Design N Population Outcome Can deep learning judge how good a predicted protein structure is? Supports Computational / modellingDeep belief network (two RBM layers plus logistic output) trained on CASP8-10, 3DRobot and native structures with five-fold cross-validation, blind-tested on CASP11 stage 1 and 2 models and an ab initio decoy set. N=84 · 84 CASP11 protein targets for the main blind test; training and testing together used 803 proteins with 216,875 structural models; ab initio validation used 24 targets. Predicted 3D protein structure models from CASP and decoy sets Per-target Pearson correlation between predicted and true GDT-TS quality, and per-target loss of the top-ranked model Can an equivariant graph network judge protein models better? Supports Computational / modellingNew equivariant graph neural network (EnQA) trained on CASP and AlphaFold2 structural models and compared with existing quality-assessment methods on held-out test sets No single N: separate test sets of 9501 CASP14 models (64 targets), 945 CAMEO models (38 targets) and 56 AlphaFold2 test targets Predicted 3D protein structural models from CASP, CAMEO and AlphaFold2 Agreement (MSE, MAE, Pearson correlation) between predicted and true lDDT quality scores at residue and model level Accurate end-point prediction does not mean the model learned folding physics.
Structure predictors including AlphaFold2 and RoseTTAFold did not predict folding type or folding rates better than chain length alone, and their predicted intermediates were mostly no better than random against experiments.
Debates
Tensions and limits
Some items are genuine disagreements on the same question. Others mark different assays, populations, or outcomes.
More precise contacts did not always give better 3D models: MetaPSICOV's more precise stage two produced slightly worse models than stage one because its extra contacts were redundant neighbours in beta sheets.
More precise contacts did not always give better 3D models: MetaPSICOV's more precise stage two produced slightly worse models than stage one because its extra contacts were redundant neighbours in beta sheets.
Qualified studies asking the same question reach different answers. The disagreement is listed, not scored.
Timeline
How understanding moved
Study years are when the paper was published. Evidence edits are dated changes to this page's claims. Explanations are when PaperFren added a Discovery — not a claim that the science happened that day.
2026
- Protein structure predictors get the answer without learning how proteins fold
Concept page published
Machine learning for protein structure prediction
Change log
What changed
Dated edits to this page's evidence: studies added or removed from a claim, claims added or withdrawn, and new explanations tagged here. Rewordings are not listed.
- Concept page published
- Protein structure predictors get the answer without learning how proteins foldEvidence: Preliminary
Papers
6 studies in this library bear on Machine learning for protein structure prediction, ordered by citations.
- Does combining predictors find more protein contacts?
A neural network that learns when to trust three different evolution-based contact predictors, and when to fall back on simpler structural clues, predicts which parts of a protein touch far more accurately than any one method.
- Can an SVM predict which amino acids touch in a folded protein?
A support vector machine using many sequence-derived features predicted residue contacts better than the previous leading method, though accuracy remained low overall.
- Can deep learning judge how good a predicted protein structure is?
A deep belief network that combines existing quality scores ranked predicted protein structures as well as or better than the best single-model methods of its time.
- Which patterns in protein sequence data reveal 3D contacts?
The 'weak' low-variance directions in protein sequence correlations — the ones principal component analysis throws away — turn out to carry most of the information about which residues touch in 3D.
- Do protein structure predictors understand how proteins fold?
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.
- Can an equivariant graph network judge protein models better?
A graph network built to respect 3D rotations and translations estimated the quality of predicted protein structures more accurately than earlier methods, and even better than AlphaFold2's own confidence scores.
Compare studies
Select 2–10 studies. Design and N are labels, not a ranking.
Nothing selected yet.
Questions
What is still open
More precise contacts did not always give better 3D models: MetaPSICOV's more precise stage two produced slightly worse models than stage one because its extra contacts were redundant neighbours in beta sheets.
Ask PaperFren about Machine learning for protein structure prediction
Study this conceptflashcards and short-answer questions
Trace how contact prediction accuracy improved from SVMs to coevolution meta-predictors.
SVMcon (2007) used hand-built features and reached about 27% accuracy at break-even, mostly wrong at long range. Coevolution methods like DCA and its reduced Hopfield-Potts version used correlated mutations in alignments, finding contacts with high accuracy in well-populated families. MetaPSICOV (2015) combined several coevolution predictors with a neural network, gaining 38% precision over the best single method, while noting that higher precision did not always improve 3D models.
Why is model-quality assessment needed even with highly accurate predictors?
Users need to know which predicted models or regions to trust. An equivariant GNN found that AlphaFold2's self-score systematically overestimated quality, while its own estimates did not differ significantly from the truth; the test set was small (56 targets), and the authors note some apparent overestimation may reflect errors in reference structures.
Flashcards
0 of 8 answers reviewed