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.
Source
From principal component to direct coupling analysis of coevolution in proteins: low-eigenvalue modes are needed for structure prediction
Study at a glance
- Design
- Computational / modelling — Statistical-physics (maximum-entropy) inference of Hopfield-Potts patterns from multiple sequence alignments, evaluated by residue-contact prediction against crystal structures while varying the number and type of patterns and the alignment size.
- N
- No single N: three protein families analysed in detail, with checks on 15 further families; contact accuracy measured on the top-ranked residue pairs per family.
- Population
- Multiple sequence alignments of protein domain families (e.g. Kunitz/BPTI, response regulator, Ras)
- Outcome
- Fraction of predicted residue-residue contacts that are true contacts in the known 3D structure
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors modelled aligned sequences of a protein family with an inverse Hopfield-Potts model, whose 'patterns' come from eigenvectors of the residue correlation matrix. Keeping all patterns reproduces direct coupling analysis (DCA); keeping only the largest ones resembles principal component analysis (PCA). They ranked patterns by their contribution to the model's likelihood and tested how well different subsets predicted residue contacts in known crystal structures, in three families in detail and 15 more as a check, and also on artificially shrunken alignments.
What they found
Contact predictions from the Hopfield-Potts model with a reduced set of patterns were essentially as good as full DCA; for the trypsin inhibitor family 96% of the top predicted contacts were real. Repulsive patterns (small eigenvalues) gave nearly all the contact accuracy, whereas using only attractive, PCA-like patterns sharply reduced it. With very small alignments of 10–30 sequences, the reduced model still found contacts with 70–80% accuracy while DCA fell to about 30%.
The limits
What it doesn't show
On large alignments the method matched DCA but did not beat it, so its practical advantage is mainly for small alignments and speed. The number of patterns to keep was not determined from sequence data alone, and the authors note that phylogenetic dependence between sequences and a heuristic pseudo-count are handled only roughly. Evaluation is limited to contact accuracy in a modest set of families, not full structure prediction.
Key terms
- Principal component analysis (PCA)
- Finding the directions (eigenvectors) of largest variance in data, usually keeping only the top ones.
- Direct coupling analysis (DCA)
- A maximum-entropy model that separates direct pairwise couplings between sequence positions from indirect correlations passed along chains.
- Maximum entropy model
- The least-assuming probability model that still reproduces chosen statistics of the data, such as pairwise correlations.
- Hopfield-Potts model
- A model whose couplings are built from a small number of patterns, here used to interpolate between PCA and DCA.
- Repulsive pattern
- A low-eigenvalue direction in sequence space that real sequences avoid; here these localise on residue pairs in contact.
- Overfitting
- Fitting noise in limited data because a model has too many parameters, which hurts predictions.
Flashcards
0 of 9 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
Which patterns carried most of the residue-contact information?
Common questions
Why does DCA struggle with small alignments?
It has a very large number of coupling parameters relative to the number of sequences, so it overfits; keeping only a few patterns cuts parameters and noise.
Why would low-variance directions matter for contacts?
Repulsive patterns concentrate on a few position pairs and penalise sequences where only one partner of a covarying pair mutates, which is exactly the signature of residues in contact.
Is this a deep learning method?
No — it is a statistical-physics inference method, but it illustrates core ML ideas: regularising by dimensionality reduction and that the most variable directions are not always the informative ones.
More on Protein structure prediction