Protein interaction networks are the scaffold on which most graph-based methods in cancer genomics run, and a large share of their edges are experimental artifacts. The spurious links bridge unrelated pathways and turn modular networks into dense "hairballs." Graph neural networks (GNNs) compute each gene's representation by repeatedly averaging over its network neighbors. On such a scaffold they fail in two linked ways. They pull signal across false bridges, and when stacked deeply they "over-smooth": every node's representation drifts toward the same vector until genes can no longer be told apart. The usual fix for the first problem is graph attention, which learns to down-weight some neighbors. That fix provably cannot address the second. The softmax attention weight is never exactly zero, so it never disconnects anything, and representations inside a connected region still collapse at an exponential rate (Demystifying Oversmoothing in Attention-Based Graph Neural Networks).
We propose to learn a per-edge gate that can switch interaction edges fully off. The gate is informed by each edge's experimental support and by the disease-specific features of the two genes it connects. Every gene keeps a protected backbone of its most reliable links, so no gene is cut off. The claim is that noise and over-smoothing on interactomes are largely one problem. Averaging within a true functional module is the useful denoising that network biologists rely on. Averaging across spurious cross-module bridges is what drives harmful collapse. A gate that cuts bridges but keeps modules should therefore both clean the graph and let depth help rather than hurt.
Core Claim
Two literatures disagree about smoothing. Network biology treats neighborhood averaging as the reason continuous embeddings tolerate noisy edges: a spurious link gets outvoted by consensus (To Embed or Not: Network Embedding as a Paradigm in Computational Biology). Deep graph learning treats the same averaging as a disease. A recent position paper argues that smoothing is harmless when labels are themselves smooth over the graph (Position: Don't be Afraid of Over-Smoothing And Over-Squashing). Our reading resolves the conflict in a way that can be tested. Cancer-driver status is roughly smooth within pathways and not smooth across the bridges that assay noise adds. Wu et al.'s theorem says collapse happens within connected components. So if the only edges set to zero are cross-module false positives, collapse becomes collapse within each module, which is roughly what the task needs. The measurable prediction concerns effective rank, the entropy of the singular-value spectrum of the node-representation matrix, which Are we measuring oversmoothing in graph neural networks correctly? found tracks accuracy with an average Pearson correlation of 0.91 where Dirichlet energy does not. On interactomes, the per-layer decay of effective rank should scale with the fraction of spurious cross-module edges. Exact-zero gating should flatten that decay, and softmax attention should not.
The closest method is DYNAMO-GAT (A Dynamical Systems-Inspired Pruning Strategy for Addressing Oversmoothing in Graph Attention Networks), which also cuts attention edges to stop collapse. Its pruning rule points the opposite way from what biological denoising needs. DYNAMO-GAT removes edges between nodes whose features are becoming correlated. In an interactome, those are disproportionately true within-module edges. A spurious bridge usually joins genes whose features disagree. DYNAMO-GAT also looks only at learned features, ignores whatever the experiments say about each edge, and its authors note that it may fragment sparse graphs. Disease pathways are exactly such sparse graphs, since they are under-sampled rather than over-connected (An integrative network-based approach to identify novel disease genes and pathways). Our gate cuts low-agreement, low-support edges, protects a backbone, and treats reliability as an input rather than something inferred from representation drift.
Research Program
The model is a deep GNN over a fixed interaction graph. Each edge $(i,j)$ carries a gate $g_{ij} \in [0,1]$ drawn from a hard-concrete distribution, a relaxation of a binary variable that places real probability mass on exactly zero. The gate enters message passing as
$$\mathbf{h}_i^{(l+1)} = \sigma\Big(\sum_{j \in \mathcal{N}_i} \frac{g_{ij}\, a_{ij}^{(l)}}{\sum_{k \in \mathcal{N}_i} g_{ik}\, a_{ik}^{(l)}}\, \mathbf{W}^{(l)} \mathbf{h}_j^{(l)}\Big),$$
where $a_{ij}^{(l)}$ is an ordinary positive attention score as in Graph Attention Networks, $\mathbf{W}^{(l)}$ is the layer's weight matrix, and $\mathcal{N}_i$ is the neighborhood of gene $i$. Attention keeps reweighting smoothly, and the gate supplies the ability to disconnect that attention lacks.
The gate's logit comes from a small network with three kinds of input. The first is the edge's static evidence: its database confidence, the number of independent evidence types supporting it, and sequence similarity between the two proteins. Zhou et al. found sequence similarity to be the most useful static signal for adding edges to protein interaction networks (Protein Function Prediction Based on PPI Networks: Network Reconstruction vs Edge Enrichment). The second is the disagreement between the two genes' input omics features, such as mutation and expression profiles. The third is a topological reliability score in the spirit of Modelling Self-Organization in Complex Networks Via a Brain-Inspired Network Automata Theory Improves Link Reliability in Protein Interactomes. The gate reads input features rather than layer-wise hidden states. This avoids the feedback loop in which already-smoothed features decide what to smooth next. It also means one learned graph is shared across all layers, which is the "change the connectivity" intervention that Wu et al.'s analysis says is required.
Two constraints encode the negative results in the file. First, for each gene the single edge with the highest static support is never gated, so no gene is disconnected. Zhou et al. showed that discarding experimental edges wholesale loses signal that similarity cannot recover, and DYNAMO-GAT's authors flag disconnection as the failure mode in sparse graphs. Second, an L0 penalty on the expected number of open gates sets how aggressively edges are cut. The gate network is trained per disease cohort, so the cut graph can differ between cancer types. A static filter cannot do that.
The analysis side reuses the rank diagnostics of Zhang et al. We record effective rank and numerical rank at every layer, and also rank within each annotated pathway versus rank across pathways. That split distinguishes good smoothing (collapse within modules, spread across them) from bad smoothing (global collapse). Dirichlet energy is recorded for contrast, because the energy-style stabilizers used in ECD-CDGI have never been checked against rank collapse.
Baselines
The naive reference is a GCN and a GAT at matched depths on the raw interaction graph, with residual connections. ECD-CDGI's ablation found that removing residual connections hurt most, so omitting them would make the baseline artificially weak.
The strongest existing task model is ECD-CDGI, which combines energy-constrained diffusion with attention for cancer-driver identification on the PPNet, PathNet and GGNet graphs. On PPNet it reports an AUC of 86.24% and an AUPR of 85.29% (ECD-CDGI). No reported numbers were found on Lacuna for its PathNet or GGNet results, or for its individual baselines. We will rerun the released configuration on identical splits rather than rely only on the quoted figure. Beating it means a higher AUPR on PPNet across repeated splits, with the gap exceeding split-to-split variance. We also expect the gap to widen under injected noise.
The strongest mechanism competitor is DYNAMO-GAT, run at the same depths with its authors' pruning schedule. On ogbn-products it reports 75.3% accuracy against 73.9% for G2GAT, and an accuracy-per-GFLOP ratio of 137.53 against 35.11 for a standard GAT on Cora (DYNAMO-GAT). It has no reported result on biological networks. For it to count as beaten, our model should score higher AUPR and should also remove a lower fraction of planted true within-module edges.
The static competitor is the best confidence-threshold filter. We sweep a threshold on database confidence, with sequence-similarity enrichment as in Zhou et al., choose it on the same validation split, and train the same deep GNN on the result. Two other graph-editing baselines complete the set: PairNorm (PairNorm) as a pure anti-collapse normalizer, and spectral pruning (Spectral Graph Pruning Against Over-Squashing and Over-Smoothing).
Three ablations isolate the key ingredient. The first replaces the hard-concrete gate with a gate that never reaches zero, which tests whether exact zeros matter. The second drops the static-evidence inputs, which tests whether the prior matters. The third drops the protected backbone, which tests the sparsity risk. Every method gets the same hyperparameter search budget, the same number of seeds and the same GPU hours. Every method is tuned on validation only.
Why not just…?
Why not just threshold the database confidence score and train a normal GNN? This is the most serious threat, and we treat it as a result that could turn out to be true. A static filter cannot adapt to a particular cancer cohort, and it cannot keep a low-confidence edge that the omics features support. Zhou et al. also show that static edge removal loses signal. Still, if the tuned threshold matches the learned gate on AUPR and rank profiles, the honest conclusion is that exact-zero pruning matters and learning does not. We will report it that way. The gate receives the confidence score as an input, so at worst it should match the threshold. The decisive comparison is the per-cohort experiment. We measure whether the learned cut graphs differ meaningfully between cancer types, and whether swapping one cohort's graph into another cohort's model lowers accuracy.
Why not just use DYNAMO-GAT, which already prunes to stop over-smoothing? Its rule removes edges between nodes whose features are becoming correlated, and in a modular interactome those are mostly true edges. The planted-noise experiment below tests this directly by measuring which kinds of edges each method removes. If DYNAMO-GAT turns out to remove mostly planted bridges anyway, our mechanism story is wrong. It would then be the method to beat, and we would say so.
Why not just stay shallow with residual connections, since driver prediction may not need depth? This is plausible. ECD-CDGI does well without extreme depth. We therefore do not claim that depth is essential. The first claim is that gating improves the best model at any depth, and the second is that it moves the optimal depth deeper. If the best gated model is still a two-layer model, the contribution reduces to denoising, and the rank analysis will show whether shallow models are simply avoiding the cross-bridge collapse that gating removes.
Why not just use sparse attention such as sparsemax without any biological prior? Sparse attention computed from features inherits the known weakness that noisy input attributes corrupt the weights. The ablation that removes static-evidence inputs from the gate measures this exactly.
Proposed Experiments
The first experiment is controlled. Starting from the PPNet interactome with Network of Cancer Genes driver labels, we treat edges supported by multiple independent evidence types as a trusted core. We then plant false positives at rates from 0% to 50% of the edge count. The planted edges are hub-biased, mimicking the degree distortions described by Zou et al.. We also delete up to 30% of true edges. For GCN, GAT, DYNAMO-GAT, the tuned threshold and our gate at depths 2, 8, 32 and 64, we report AUPR, per-layer effective rank split into within-pathway and across-pathway components, and the precision and recall of each method's removals against the planted edges. Our hypothesis fails if GAT's rank decay does not steepen with the planted-bridge fraction, or if gating does not flatten it.
The second experiment is the real benchmark: driver identification on PPNet, PathNet and GGNet against the ECD-CDGI figures, plus a transfer check on BioGRID and STRING function prediction in yeast and mouse following Zhou et al.'s protocol. Those datasets have no reported numbers on Lacuna, so we rerun that baseline ourselves. A Cora experiment with planted noise tests whether the effect is specific to modular biological noise or general to any graph.
The decisive checks are three. The learned gate should beat the tuned confidence threshold under realistic noise levels. Across cohorts, the gate should cut different edges, and those differences should matter for accuracy. And at depth, within-module rank should fall while across-module rank stays high. That last pattern is the direct signature that smoothing has become denoising rather than collapse.
References
- Demystifying Oversmoothing in Attention-Based Graph Neural Networks
- To Embed or Not: Network Embedding as a Paradigm in Computational Biology
- Position: Don't be Afraid of Over-Smoothing And Over-Squashing
- Are we measuring oversmoothing in graph neural networks correctly?
- A Dynamical Systems-Inspired Pruning Strategy for Addressing Oversmoothing in Graph Attention Networks
- An integrative network-based approach to identify novel disease genes and pathways: a case study in the context of inflammatory bowel disease
- Graph Attention Networks
- Protein Function Prediction Based on PPI Networks: Network Reconstruction vs Edge Enrichment
- Modelling Self-Organization in Complex Networks Via a Brain-Inspired Network Automata Theory Improves Link Reliability in Protein Interactomes
- ECD-CDGI: An efficient energy-constrained diffusion model for cancer driver gene identification
- PairNorm: Tackling Oversmoothing in GNNs
- Spectral Graph Pruning Against Over-Squashing and Over-Smoothing
- Systematic Analysis of Pleiotropy in C. elegans Early Embryogenesis