Interpretability
Can wiring a neural net by biology help small-data prognosis?
Open access · cc by · source: Europe PMC
A neural network whose first connections follow known gene pathways, and which prunes itself to stay sparse, predicted brain-tumour survival slightly better than standard classifiers while being easier to interpret.
Study at a glance
- Design
- Computational / modelling — New pathway-masked sparse neural network vs four baseline classifiers, stratified 5-fold cross-validation repeated ten times on TCGA glioblastoma gene expression.
- N
- N=475 · Patient samples after excluding censored cases: 99 long-term survivors (past 24 months) and 376 short-term survivors; 4359 genes in 574 Reactome pathways as inputs.
- Population
- Glioblastoma patients' tumour gene-expression profiles from The Cancer Genome Atlas
- Outcome
- Classification of long-term vs short-term survival, measured by AUC and F1-score
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
PASNet had the highest mean AUC (0.6622) and F1-score (0.3978) across the repeated cross-validation, ahead of the dropout network (AUC 0.6408) and SVM (0.6337), and the authors report these differences as statistically significant. Nonlinear models beat the linear LASSO models. Inspecting the trained weights highlighted pathways such as GPCR signalling and the innate immune system, many of which the authors note have prior links to glioblastoma.
Methodology
The authors built PASNet, a neural network with a gene input layer, a pathway layer connected only to each pathway's member genes, one hidden layer and an output layer. It was trained by repeatedly training small dropout sub-networks, pruning weak weights at a per-layer sparsity level, and weighting errors by class to handle imbalance. Using glioblastoma gene expression from TCGA, they predicted whether patients survived beyond 24 months and compared PASNet with an SVM, random LASSO, LASSO logistic regression and a same-size fully connected dropout network.
Limitations
Absolute accuracy is modest: an AUC of 0.6622 means the model separates survivors only somewhat better than chance, and the gains over baselines are a couple of points. Everything was evaluated on one cancer dataset with no external validation cohort, and hyperparameters were tuned separately for PASNet and the dense network by hand-guided search. The biological interpretation is post hoc; the authors themselves caution that a node being active in the network does not mean the pathway is biologically activated.
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.
Knowledge-structured networks give readable parts, with modest gains.
Building biology into the architecture can make networks both usable and somewhat more accurate: PASNet's pathway-constrained network had the highest mean AUC (0.6622) for glioblastoma prognosis, and TranSynergy improved drug-synergy correlations by 3–5% over DeepSynergy.
Evidence for the claim as stated.
Knowledge-structured networks give readable parts, with modest gains.
Building biology into the architecture can make networks both usable and somewhat more accurate: PASNet's pathway-constrained network had the highest mean AUC (0.6622) for glioblastoma prognosis, and TranSynergy improved drug-synergy correlations by 3–5% over DeepSynergy.
Scope note — AUC 0.66 is modest, one dataset, no external validation.
Limits the claim's scope: a different population, assay, or outcome.
An explanation is a lead to check, not a finding.
Explanations are post hoc hypotheses unless tested: PASNet's authors caution active nodes don't mean activated pathways, TranSynergy's SHAP explanations were not validated in the lab, and the EHR model's attribution example was a single case.
Evidence for the claim as stated.
'Interpretable' means different things: point scores are interpretable by simplicity (not measured with clinicians), knowledge-guided networks by structure, and a robot navigation paper uses 'explainable' to mean reward terms have physical meaning, not that the policy was analysed.
Evidence for the claim as stated.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
'Interpretable' means different things: point scores are interpretable by simplicity (not measured with clinicians), knowledge-guided networks by structure, and a robot navigation paper uses 'explainable' to mean reward terms have physical meaning, not that the policy was analysed.
Related papers in this topic
Same topic cluster — not a recommendation engine.