Can SVMs predict how membrane proteins sit in the membrane?
A set of support vector machines predicted the full membrane topology correctly for 89% of test proteins, beating earlier methods, though re-entrant helices remained hard.
Source
Transmembrane protein topology prediction using support vector machines
Study at a glance
- Design
- Computational / modelling — Four SVMs on evolutionary profiles combined by dynamic programming, plus a TM-vs-globular SVM, fully cross-validated with homologues removed
- N
- N=131 · 131 transmembrane proteins with crystal structures used for cross-validated testing
- Population
- Alpha-helical transmembrane protein sequences
- Outcome
- Per-residue Matthews correlation and percentage of proteins with fully correct topology
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors assembled a new dataset of transmembrane proteins whose topology comes only from crystal structures. They trained separate SVMs to label residues as membrane helix or not, inside or outside loop, signal peptide, and re-entrant helix, using evolutionary profiles, then combined outputs with dynamic programming into ranked topologies. They cross-validated with homologous proteins removed and compared with other predictors, and trained an extra SVM to tell membrane from globular proteins.
What they found
The method predicted the complete topology for 116 of 131 proteins (89%), versus 79% for the next best method, and got the number of helices right 95% of the time. Different SVMs needed different kernels: the loop SVM reached an MCC of 0.63 with a polynomial kernel but only 0.35 with an RBF kernel. Signal-peptide proteins were handled well (13 of 14), but re-entrant helix proteins were not (7 of 11), mainly because there were few training examples.
The limits
What it doesn't show
Competing methods were run from web servers without cross-validation, and one was trained on 92% of the test proteins, so comparisons are not perfectly fair (the authors note this likely inflates rivals). Subgroups like signal-peptide and re-entrant proteins are tiny, so percentage differences can reflect a single protein. The dataset is limited to structurally solved proteins, which may not represent all membrane proteins in genomes.
Key terms
- Support vector machine
- A classifier that finds the boundary separating classes with the widest margin, optionally via a kernel.
- Kernel function
- A similarity function that lets an SVM draw non-linear boundaries, e.g. polynomial or radial basis function.
- Membrane protein topology
- The number, positions and in/out orientation of a protein's membrane-spanning helices.
- Matthews correlation coefficient
- A balanced measure of binary classification quality from -1 to 1, robust to class imbalance.
- Cross-validation
- Repeatedly training on part of the data and testing on the rest, so every item is tested by a model that never saw it.
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
What proportion of test proteins had fully correct topology?
Common questions
Why use several SVMs instead of one?
Each residue type is best separated by a different kernel; for example the loop classifier worked far better with a polynomial kernel than an RBF one.
Why remove homologous proteins from training?
Similar sequences in training and test sets would let the model memorise answers and overstate accuracy.
Why are re-entrant helices hard to predict?
They look like membrane helices and there were very few independent training examples.
More on Protein structure prediction