Many cancer survival models built on gene expression wire their hidden layer to match known biology. Each hidden unit is connected only to the genes of one documented pathway, such as a KEGG or Reactome signaling cascade, and the model's explanation is read off these "pathway neurons." The unit is only called a pathway, though. Nothing checks that its activation measures that pathway's activity rather than an arbitrary weighted mix of its genes. That mix can be whatever best predicts survival in a cohort of a hundred patients. Concept bottleneck research on images has shown that this kind of gap is real and harmful: continuous intermediate units carry hidden label information ("leakage"). Explanations built on them are then unfaithful, and a human who corrects an intermediate value can make predictions worse.
We propose to test whether pathway-masked survival networks leak in the same way, and to build a version that cannot. In this model, each pathway unit is tied to a label-free activity score learned without survival data. The unit may deviate from that score only through a small, capped residual. The final hazard is an additive function of pathway units that stays linear unless the data demand otherwise. The claim is falsifiable. In small TCGA cohorts, much of the accuracy advantage of pathway-masked networks over fixed pathway scores should come from leakage that anchoring removes. The anchored model should keep most of the accuracy, and its pathway units should become things one can intervene on.
Core Claim
Two things need separating. The literature treats the hidden layer of a "visible neural network" (a network whose connectivity copies a biological graph) as interpretable because of how it is wired. The design principle is the Hadamard mask $h=\sigma((W\odot M)x+b)$, where $M$ zeroes every gene-to-pathway weight that the database does not list. The critical review of these models (Visible neural networks for multi-omics integration: a critical review) documents large parameter savings and competitive accuracy. PathExpSurv (PathExpSurv), consDeepSignaling (Predicting Anticancer Drug Response With Deep Learning Constrained by Signaling Pathways) and PONET (Deep Biological Pathway Informed Pathology-Genomic Multimodal Survival Prediction) all read explanations off such units.
Wiring constrains which genes a unit may see. It does not constrain what the unit computes from them. A Reactome pathway with 80 genes gives a trainable projection with 80 free weights. When the number of patients is small and the number of genes is large, that projection can encode survival-relevant noise that has nothing to do with the pathway's biological state. This is the same failure that Addressing Leakage in Concept Bottleneck Models and Hi-CBM proved for soft concept bottlenecks. In those models, continuous concept probabilities smuggle class information, and test-time corrections then fail. Pathway networks have never been audited for it.
The central claim has three parts. Pathway units in masked survival networks leak measurably in small cohorts. A random mask with the same number of connections per unit should recover much of their accuracy. Anchoring each unit to an independently learned activity score removes the leak and costs little concordance.
Research Program
The work has two halves: an audit of existing pathway networks, and a model designed to pass that audit.
The audit. We define two quantities for any pathway unit $z_p$ (the activation of pathway $p$).
The first is anchor agreement: the correlation of $z_p$, in held-out patients, with a label-free activity score $a_p$ for the same pathway. The score $a_p$ comes from an unsupervised pathway autoencoder in the style of NetActivity, trained on expression data with the survival labels of the target cohort withheld.
The second is residual leakage: the concordance index (C-index, the probability that the model ranks a random comparable patient pair in the right survival order) of a Cox model fit only on the residuals $z_p - \hat{E}[z_p\mid a_p]$. High residual leakage together with low anchor agreement means the unit is a survival feature under a pathway's name.
We will also run the intervention test that the concept bottleneck literature uses. We corrupt the input genes of a chosen set of pathways, simulating dropout or batch noise, and then "correct" those units by setting them to their anchor values computed from clean data. If the units are faithful, predictions recover toward those from uncorrupted inputs. If they leak, predictions do not recover and may get worse. That is the analogue of the degradation reported in Addressing Leakage in Concept Bottleneck Models.
The anchored model. Each pathway unit is written as a frozen anchor plus a capped residual:
$$z_p(x) = a_p(x) + r_p(x),\qquad \sum_p \lVert r_p \rVert^2 \le \tau,$$
where $x$ is a patient's expression vector, $a_p$ is the frozen label-free activity score, $r_p$ is a masked, trainable correction restricted to pathway $p$'s genes plus a sparse set of added genes, and $\tau$ is a residual budget chosen on validation data. The added genes follow PathExpSurv's L1 expansion but are confined to the residual.
The budget turns PathExpSurv's open-ended expansion into a quantity we can report. Every added gene can shift a pathway unit only within a stated amount, so the unit cannot drift into a dense, opaque projection. That drift is the unresolved risk of both PathExpSurv's expansion and SENA's leakage parameter $\lambda$ (Interpretable Causal Representation Learning for Biological Data in the Pathway Space).
Residual variance that no pathway explains goes into a single explicit side channel $s(x)$ with its own reported contribution. This follows the latent side channel of the autoregressive hard bottleneck in Addressing Leakage in Concept Bottleneck Models, which is honest about incompleteness instead of hiding it inside pathway units.
The hazard is an additive model over pathway units. Each term starts linear and gains a spline only when a concordance-based criterion admits it. This is the "reluctant" principle of RAITR, which favors sparse linear rules and fits nonlinear terms only if they add clear accuracy under a stronger penalty. Here we transfer it from genes to pathway units.
We deliberately do not use a fixed binary pathway-to-class matrix as in Hi-CBM. That paper's own limitation is that graded prognostic risk does not fit deterministic concept-to-class maps.
Scope. We treat multimodal fusion with slide images as out of scope for the core claim. PONET's cross-attention between images and pathways is a natural follow-up once the unimodal pathway units are shown to be faithful. Fusing images into leaky units would only compound the problem.
Baselines
The naive reference point is a Cox model with elastic-net penalty fit directly on genes. On TCGA-KIRP, PONET reports a standard Cox C-index of 0.544 (PONET). We will refit it in our own splits rather than reuse that number.
The most dangerous baseline is the one a skeptic would choose. It computes the same frozen anchor scores $a_p$ and fits a sparse Cox model on them. By construction it is interpretable and cannot leak. If it matches every pathway network, then pathway networks buy nothing, and our anchored model reduces to it plus decoration.
The strongest existing pathway network in this setting is PathExpSurv. It reports higher C-index than Cox, Random Survival Forests and DeepOmix on BRCA, LGG and THCA, and beats a dense network on THCA. However, no reported result was found on Lacuna for PathExpSurv's numeric C-index. PONET's 0.829 on KIRP is the only strong pathway-network number in the file, and it uses slides plus three omics types. It is therefore an upper reference and not a like-for-like comparison.
We also include three further comparators. Random Survival Forests with TreeSHAP represent the tree-ensemble alternative. A dense multilayer perceptron with SHAP represents the black-box-plus-attribution alternative, which consDeepSignaling reports at Pearson about 0.74 against its constrained model's about 0.85 on GDSC (consDeepSignaling). A pathway network with degree-matched random masks tests whether database identity matters at all.
Three ablations isolate the anchored model's key ingredients: removing the anchor (a plain masked unit), removing the budget (an uncapped residual), and removing reluctance (all-nonlinear terms).
Fairness rules apply to every method. All methods see identical patient splits, the same gene filtering, and the same pathway database. Each gets the same number of hyperparameter trials, tuned on inner validation folds only, and anchors are never trained on target-cohort labels.
Beating the field means three things at once. The anchored model must come within a small, pre-registered C-index margin of PathExpSurv, with the margin fixed before seeing results. It must beat the anchored-score Cox model by a margin whose confidence interval excludes zero across subsampled sizes. It must also show clearly lower residual leakage and better intervention recovery than PathExpSurv and the unanchored ablation.
Why not just…?
Why not just fit a sparse Cox model on pathway activity scores? This is the real threat, and we treat it as the null hypothesis. The file's evidence against decoupled pipelines (consDeepSignaling; PONET) comes from larger or multimodal settings. It does not come from small unimodal survival cohorts. If the anchored model's advantage over this baseline vanishes at small cohort sizes, we will report that anchored scores plus a linear Cox model are sufficient. We would then treat the leakage audit of existing networks as the main contribution.
Why not just use PathExpSurv and read its pathway weights? Its expansion phase lets pathway units absorb any correlated genes. Its bootstrapping reports which links are stable, but not whether a unit still measures its pathway. The residual-leakage and intervention tests measure exactly this. If PathExpSurv shows low leakage, our motivation weakens and we will say so.
Why not a dense network with SHAP, or a random forest with TreeSHAP? Both can be accurate. Their attributions, however, describe genes, not intermediate states one can set and re-run. The concept bottleneck literature (Integrating Clinical Knowledge into Concept Bottleneck Models) shows that post-hoc attributions track shortcuts that fail out of domain. Our cross-cohort transfer experiment tests whether that failure appears here.
Proposed Experiments
We use TCGA RNA-seq with overall survival for BRCA, LGG, THCA and KIRP, the cohorts used by PathExpSurv and PONET. Pathways come from Reactome and KEGG. We subsample training sets to 50, 100 and 200 patients, with 20 repeats per size, and evaluate on fixed held-out patients using C-index and integrated Brier score. We add a transfer test that trains on one kidney cohort (KIRP or KIRC) and evaluates on the other, to probe institutional and biological shift.
Four checks are decisive. The first compares real and random masks. If random masks match real ones in C-index in existing pathway networks, pathway identity is not doing the predictive work. The second is the leakage audit. Residual leakage in unanchored units should rise as cohorts shrink, and should fall toward chance in anchored units. The third is the intervention test. Anchored units should restore predictions after input corruption, while unanchored units should not. The fourth is the accuracy cost of faithfulness, meaning the C-index gap between the anchored model and both PathExpSurv and the anchored-score Cox model at each cohort size.
Each outcome has a clear reading. If anchoring costs little accuracy and removes leakage, pathway networks can be both accurate and interventionable. If it costs a lot, their accuracy was coming from the leak.
References
- Visible neural networks for multi-omics integration: a critical review
- PathExpSurv: pathway expansion for explainable survival analysis and disease gene discovery
- Predicting Anticancer Drug Response With Deep Learning Constrained by Signaling Pathways
- Deep Biological Pathway Informed Pathology-Genomic Multimodal Survival Prediction
- Addressing Leakage in Concept Bottleneck Models
- Hi-CBM: Mitigating information leakage via hierarchical concept bottleneck modeling
- NetActivity enhances transcriptional signals by combining gene expression into robust gene set activity scores through interpretable autoencoders
- Interpretable Causal Representation Learning for Biological Data in the Pathway Space
- A reluctant additive model framework for interpretable nonlinear individualized treatment rules
- Integrating Clinical Knowledge into Concept Bottleneck Models