Can a known gene network make microarray classifiers interpretable?
Forcing a classifier's gene weights to vary smoothly across a known metabolic network kept accuracy unchanged while making the classifier readable in terms of pathways.
Source
Classification of microarray data using gene networks
Study at a glance
- Design
- Computational / modelling — Expression profiles projected onto low-frequency eigenvectors of the KEGG metabolic-network Laplacian (threshold or exponential filters), then used for unsupervised distance analysis/PCA and a leave-one-out SVM, compared with the unfiltered Euclidean space and 100 vertex-permuted random networks.
- N
- N=17 · 17 yeast expression profiles (twelve non-irradiated cultures minus one outlier, plus 6 irradiated); network of 713 genes in the largest connected component.
- Population
- Saccharomyces cerevisiae cultures with and without low-dose irradiation; KEGG metabolic gene network.
- Outcome
- Ratio of intraclass to total distances, leave-one-out misclassifications and hinge loss, and interpretability of classifier weights on the network.
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors treated a gene expression profile as a signal on a graph whose nodes are genes linked in the KEGG metabolic network, and used the graph Laplacian's eigenvectors as a kind of Fourier basis. They removed or damped the 'high-frequency' parts of each profile, which vary sharply between neighbouring genes, then compared samples and trained a support vector machine on the filtered data. They tested this on 17 yeast cultures, irradiated or not, using leave-one-out classification and comparisons against randomly shuffled networks.
What they found
Keeping only the smoothest part of the signal (discarding up to 80% of eigenvectors) made irradiated and non-irradiated samples cluster better than with random networks, but keeping too few (1-3%) did worse; the exponential filter mostly hurt. In leave-one-out SVM classification, error stayed at the baseline of 2 misclassifications until about 80% of eigenvectors were removed, so accuracy was not improved but not lost. The smooth classifier highlighted interpretable modules, such as raised oxidative phosphorylation and a switch from glycolysis to gluconeogenesis in irradiated samples, which the unconstrained classifier did not reveal.
The limits
What it doesn't show
The test uses a single, very small dataset of 17 samples, and the authors admit the accuracy gain is only slight, so the claim that network smoothing improves prediction is not demonstrated. The same two unusual irradiated samples were misclassified by both classifiers. The method treats the metabolic network as undirected and unweighted and assumes neighbouring genes are co-expressed, which is only an average trend with many exceptions. Biological interpretations of classifier weights (e.g. kinase downregulation) are suggestions, not validated findings.
Key terms
- Graph Laplacian
- A matrix built from a network's connections whose eigenvectors describe patterns that vary smoothly or sharply across the network.
- Low-pass filter
- An operation that keeps slow, smooth variation in a signal and removes fast, jagged variation, here defined over a gene network.
- Support vector machine (SVM)
- A linear classifier that finds a separating boundary with the largest margin between two classes.
- Regularisation
- Adding a penalty to training so the model prefers simpler or more plausible solutions, which helps when there are few samples and many features.
- Leave-one-out cross-validation
- Training on all samples but one, testing on the held-out one, and repeating for every sample.
Flashcards
0 of 10 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 does the network filtering remove from expression profiles?
Common questions
Why would smoothing over a network help at all?
With thousands of genes and only a handful of samples, a classifier can fit noise; assuming that genes linked in a pathway act together gives a prior that reduces the space of plausible classifiers and groups genes into interpretable modules.
Did the method make classification more accurate?
Not really. The leave-one-out error stayed at 2 misclassifications, the same as the ordinary SVM, until too many components were removed; the main benefit was interpretability.
How does this relate to other regularisers like picking a few genes?
Feature selection assumes the best classifier uses few genes; this method instead assumes the best classifier is smooth on the gene network. Both reduce complexity, but they encode different beliefs.
More on Graph neural networks