Pan-Cancer Driver Models Over-Smooth Across Tumor Types

Models that rank cancer driver genes, meaning the genes whose alterations actually cause tumors, now routinely pool molecular data from many tumor types. They combine mutations, copy-number changes, DNA methylation and expression from 16 cancer types in The Cancer Genome Atlas (TCGA) with a protein-protein interaction network, and they output one score per gene. Producing one score from sixteen tumor types is an averaging step. In graph-learning terms it is extreme over-smoothing: the tendency of graph neural networks to push every node's representation toward the same vector as information is averaged across neighbors, with the tumor types playing the role of the neighbors. Biology tells us this averaging throws away real signal. The immune gene CD74 tracks better survival in most solid tumors but worse survival in gliomas (CD74 is a potential biomarker predicting the response to immune checkpoint blockade). A pooled score cannot represent that reversal.

We propose to treat the tumor-type axis as a graph, apply the tools of the over-smoothing literature to it, and measure and control how much of each gene's signal is shared across tumors versus specific to some of them. Concretely, a node is a pair (gene, tumor type). We separate a low-pass component, the part shared by similar tumor types, from a high-pass component, the part where one tumor type departs from its neighbors. A rank-based diagnostic then tells us whether a model's gene-by-tumor score matrix has collapsed to a single pan-cancer ranking. The new element is not a deeper network on the protein interaction graph. It is the recognition that the field's pooling step is itself the smoothing operator that needs to be diagnosed and undone.

Core Claim

Current pan-cancer driver predictors effectively output a rank-one gene-by-tumor-type matrix: the same ordering of genes in every tissue. That suits the standard benchmark, which scores genes against a single pooled label list, but it hides tumor-type-specific drivers. A filter along the tumor-type graph that splits shared signal from deviation can recover tumor-specific rankings while keeping pooled accuracy.

The split follows the decomposition of the identity into low-pass and high-pass parts introduced for graph convolution in Complete the Missing Half. That paper adds a high-pass "diversification" channel next to ordinary neighbor averaging and gains about 10% accuracy on graphs where neighbors differ. We reuse the decomposition but move it from the node graph to the tumor-type axis.

Write $X_g \in \mathbb{R}^{T \times d}$ for gene $g$'s omic features in each of $T$ tumor types, with $d$ features per type. Write $P$ for the row-normalized adjacency of a tumor-type similarity graph. The two channels are

$$ H^{\text{low}}_g = P^{k} X_g, \qquad H^{\text{high}}_g = X_g - P^{k} X_g , $$

where $k$ is the number of smoothing steps. Mean-pooling over all tumors is the special case of a complete graph, and it equals the limit of large $k$ on any connected graph. Existing pooled models therefore sit at the most-smoothed end of this family.

A per-gene gate $\alpha_g \in [0,1]$ mixes the two channels before a shallow protein-network encoder. The gate borrows the adaptive-aggregation idea of When Do GNNs Work, where a learned per-node gate stops aggregation early and reaches 73.3% accuracy on Cora with 1% labels versus 59.2% for a standard GCN. In our setting, a gene with a coherent pan-cancer role should learn $\alpha_g$ near 1, so that it relies on shared signal. A gene like CD74 should not.

Research Program

The measurement half comes first because it can falsify the premise cheaply. For any model, we collect its gene-by-tumor score matrix $S \in \mathbb{R}^{n \times T}$, where $n$ is the number of genes. For pooled models we get $S$ by feeding one tumor type's features at a time with the others masked. We then compute effective rank, the entropy-based count of significant singular values, and numerical rank, the squared Frobenius norm divided by the squared spectral norm.

Are we measuring oversmoothing in graph neural networks correctly? supplies these metrics. It shows that over-smoothing in trained networks is a collapse of representations toward a one-dimensional subspace. Effective rank tracks accuracy with an average Pearson correlation of 0.91, while Dirichlet energy, the classical smoothness measure, misses the collapse. That paper also constrains us: high rank does not imply usefulness, since random noise has maximal rank. We therefore never report rank without per-tissue accuracy beside it. We also track the same rank metrics layer by layer on the protein-network axis. This tests directly whether the 2–3-layer encoders used in the field already collapse, which nobody in cancer genomics has measured.

The modeling half has three parts. The tumor-type graph is built three ways, so the choice of $P$ is not doing hidden work: a k-nearest-neighbor graph on pan-cancer expression correlation, a graph grouping tumors by organ system, and the complete graph that reproduces mean pooling. The protein-network encoder is deliberately kept at two layers. We are not claiming depth helps on the interaction graph, and PairNorm reports that its anti-smoothing normalization lets models go deeper without raising peak accuracy on fully featured graphs. The output head produces a pan-cancer score from the low-pass channel and a per-tumor score from the gated combination.

The input data and pooled labels are those used by Explainable Multilayer Graph Neural Network for Cancer Gene Prediction and Towards simplified graph neural networks for identifying cancer driver genes in heterophilic networks. That means 29,446 TCGA samples across 16 types and four omic modalities, with 887 known cancer genes and 7,753 non-cancer genes. Graphs come from CPDB and STRING.

Tumor-specific evaluation needs labels that the pooled benchmark lacks. We use two sources. The first is the tumor-type annotations attached to Cancer Gene Census entries, restricted to tumor types with enough positives. The second is a label that is automatically available and independent of driver curation: the sign of each gene's per-tumor Cox survival association in TCGA, the kind of analysis that produced the CD74 glioma reversal. Reversals of this kind are the cases a rank-one model must get wrong in at least one tissue.

Relation to the closest work

Four papers apply graph networks to pan-cancer driver prediction. All of them score each gene once, from features pooled or concatenated across tumor types. SGCD (Towards simplified graph neural networks…) diagnoses heterophily on interaction networks, meaning that connected proteins often have different labels. It separates node and neighbor representations and reaches an area under the precision-recall curve (AUPRC) of 0.8027 on CPDB. EMGNN combines six interaction networks through a second-level graph network over per-network gene embeddings and reports a 7.15% AUPR gain over the earlier EMOGI model. SBM-GNN couples graph diffusion with community block models and reaches an area under the ROC curve (AUC) of 0.8948 on STRING. SEFGNN works on cell-line data and uses multi-hop aggregation. All four attack smoothing, if at all, on the protein-network axis. None treats tumor type as a structured axis or reports per-tissue performance.

Baselines

The naive reference is a single pooled ranking copied to every tumor type. We build it by taking SGCD's pooled scores and tiling them across tissues, which gives an effective rank of exactly one. Any tumor-specific gain must be measured against this. If per-tissue AUPRC from the tiled pooled ranking is already close to that of tissue-aware models, the premise fails.

The strongest existing method is SGCD, reproduced from its released setup on CPDB, together with MTGCN as reported in the same paper. The reported pooled AUPRCs are 0.8027 for SGCD and 0.8016 for MTGCN (Towards simplified graph neural networks…). On STRING we also run SBM-GNN, whose reported AUC is 0.8948 against 0.8660 for a GCN. For per-tumor driver ranking there is no reported result found on Lacuna for any of these methods.

Every model receives the same splits, the same 16-type features, the same graphs, and the same hyperparameter search budget of a fixed number of trials on the validation fold. On the pooled task, "not worse" means staying within 0.01 AUPRC of our SGCD reproduction. On per-tissue AUPRC and survival-sign accuracy, "beating it" means a gain whose 95% bootstrap confidence interval over genes excludes zero, averaged over tumor types with at least 20 annotated positives.

The competitor most likely to match us is a multitask baseline. It uses the same two-layer protein-network encoder on concatenated per-tumor features, with a separate output head per tumor type and no tumor-type graph. This tests whether per-tissue capacity alone suffices.

The ablation that isolates our key ingredient removes the high-pass channel and keeps only the low-pass one, with the gate fixed at one. A second ablation fixes the gate at 0.5 to test whether gating per gene matters. A third swaps the learned tumor graph for the complete graph.

Why not just…?

Why not just train a separate model per tumor type? That is the zero-smoothing extreme. With 887 pooled positives spread unevenly over 16 types, many types would have only a handful of tissue-annotated drivers. Scarcity of rare driver labels is already the field's central difficulty. We include per-type models as a baseline and expect them to lose badly on small types. If they do not, sharing across tumors is not needed, and that is a finding worth reporting.

Why not just add per-tumor output heads to an existing pooled model? This is the real threat, because it already gives the model room to produce a high-rank score matrix. The difference is that it shares nothing structured between related tumors and has no explicit notion of "deviation from similar tumors." The multitask baseline above settles the question directly. If it matches our model on per-tissue AUPRC and on survival-sign reversals, our contribution reduces to the diagnostic plus the observation that multitask heads suffice. We will report it that way rather than over-claim the filter.

Why not just make the protein-network encoder deeper with PairNorm or a high-pass filterbank? That is the obvious combination of the two literatures. The evidence suggests it buys little here. PairNorm's gains appear mainly when features are missing, and pooled benchmark numbers are already within 0.001–0.02 of each other (0.8027 versus 0.8016 versus 0.7848). We test this by running a deeper, PairNorm-equipped encoder on pooled features. We expect it to stay rank-one on the tumor axis no matter how well it preserves rank on the gene axis.

Proposed Experiments

The first decisive check is purely diagnostic. We compute the effective rank of the gene-by-tumor score matrix for SGCD, EMGNN-style and SBM-GNN-style models, the multitask baseline, and our model. Pooled models should sit near one. If they already show substantial rank and good per-tissue AUPRC, the over-smoothing framing is wrong and we stop there.

The second check is the main comparison. We evaluate pooled AUPRC on CPDB and STRING, per-tissue AUPRC on tumor-annotated drivers, and accuracy at predicting the sign of per-tumor survival associations. That last metric is reported separately on the subset of genes whose sign reverses across tumors, where CD74 is the motivating case and where a rank-one model is guaranteed to fail somewhere.

The third check is leave-one-tumor-type-out generalization. We hold out a tumor type's labels and features during training, then feed its features at test time. This asks whether the tumor-type graph lets a new tissue borrow from its neighbors better than per-tumor heads can.

Finally, we inspect the learned gates. Genes in the genome-stability module identified by Weighted Frequent Gene Co-expression Network Mining, 412 mitotic and DNA-repair genes recurrent across 33 cancer types, should receive gates near one, meaning they rely on shared signal. Immune and lineage genes should receive lower gates. Agreement with this independent, experimentally supported partition is a sanity check on whether the shared-versus-specific split means anything biologically.