Graph neural networks
Can a known gene network make microarray classifiers interpretable?
Open access · cc by · source: Europe PMC
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.
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.
Key findings
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.
Methodology
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.
Limitations
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.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Graph priors can aid interpretation even when accuracy does not change.
Constraining a classifier to be smooth over a metabolic gene network did not improve accuracy on a 17-sample microarray dataset but produced more interpretable pathway-level weights.
Evidence for the claim as stated.
Related papers in this topic
Same topic cluster — not a recommendation engine.
- Can spreading signals through a protein network find disease genes?
- Can a graph neural network sort unknown phage DNA into families?
- Can self-supervised learning predict how mutations change binding?
- Does letting each node choose its own depth fix GNN over-smoothing?
- Can a graph neural network judge predicted protein shapes?