Where Rare Subtypes Disappear: Imbalance-Aware Neighborhoods for Similarity Network Fusion

Many multi-omics cancer classifiers start by building a "patient similarity network", a graph in which each patient is a node and edges connect patients whose molecular profiles look alike. Each data type (gene expression, DNA methylation, microRNA) gets its own graph, and the graphs are merged into one. The standard merging method, similarity network fusion (SNF), keeps only each patient's $k$ most similar neighbors in every data type. It then repeatedly passes similarity information between the per-type graphs, so that edges supported by several data types grow stronger and edges seen in only one type fade. A graph neural network is then trained on the fused graph. This note argues that when one cancer subtype is rare, the damage is done before any classifier is trained. A rare-subtype patient whose class has fewer than $k$ close members is forced to take majority-class neighbors. Cross-type reinforcement then treats the rare subtype's distinctive signal as single-type noise and erases it. Reweighted losses applied downstream cannot recover edges the graph never kept.

We propose to measure this effect directly, as the fraction of a rare subtype's fused-graph edge weight that stays within that subtype, and to fix it at its source. The fix replaces SNF's uniform top-$k$ truncation with a per-patient neighborhood whose size adapts to local density and requires mutual agreement between neighbors, built without labels so it can be used transductively. The question that matters is whether this graph-level fix improves rare-subtype recall beyond what imbalance-aware losses, a smaller global $k$, or class-weighted gradient boosting already achieve. The experiments are designed so that a negative answer is as informative as a positive one.

Core Claim

The fixed neighborhood size in SNF's local similarity matrix is the main reason fused patient graphs fail rare subtypes. If the neighborhood is made density-adaptive and mutual, rare-subtype patients keep their within-class edges through fusion. Downstream classifiers then gain macro-F1 and rare-class recall that cannot be matched by changing only the loss or only the classifier.

The prior work that supplies the method is SNF's cross-diffusion. Each data type's local top-$k$ matrix is used to filter another type's global similarity matrix, on the stated principle that local structure is reliable and global structure is noisy (Tian et al., Wei et al.). A benchmark of ten integration methods across nine TCGA cancers found network-based fusion (SNF, NEMO) the most robust to noise (Duan et al.), so we keep cross-diffusion rather than discard it. The work that motivates the proposal comes from graph learning. DPGNN showed that on skewed single graphs, message passing smooths minority nodes into their majority neighbors. On Cora at a 1:10 imbalance it raised macro-F1 from 0.5205 (plain GCN) to 0.7115 using class prototypes. DPGNN repairs this inside the network, however, and never looks at how the graph was built from multiple views.

The closest systems all take the fused or per-type graph as fixed and train with standard cross-entropy. DeepMoIC feeds an SNF graph into a GCN of up to 20 layers and reports 84.28% overall accuracy on TCGA Pan-cancer with no per-class metrics. MOGONET uses per-type graphs and fuses class probabilities late. MOGAT reweights neighbors with attention, which can only reweight edges that already exist. None of them changes which edges exist, and that is the step we target.

Research Program

The first component is a diagnostic. For each class $c$, we define within-class retention on a graph with weights $W$ as

$$h_c(W) = \frac{\sum_{i \in c}\sum_{j \in c} W_{ij}}{\sum_{i \in c}\sum_{j} W_{ij}},$$

which is the share of class-$c$ patients' edge weight that lands on other class-$c$ patients. Labels are used only to evaluate graphs, never to build them. We will compute $h_c$ for each per-type graph, for the linear mean-fused graph, and for the SNF graph at every diffusion iteration, using TCGA-BRCA, METABRIC, and TCGA-KIPAN. The hypothesis is falsified if SNF preserves or raises $h_c$ for rare classes relative to the best single data type. To test the mechanism, we will subsample one BRCA subtype (for example Basal-like) to controlled sizes below and above $k$. If the hypothesis holds, $h_c$ should drop sharply once the class size falls under $k$.

The second component is the fix. For data type $v$ with kernel affinity $W_v$, patient $i$'s neighborhood $N_i^v$ contains the patients $j$ whose affinity exceeds a fraction $\tau$ of $i$'s strongest affinity, $W_v(i,j) \ge \tau \max_l W_v(i,l)$, and who also rank $i$ within their own top $k_{\max}$. The size of the neighborhood is clamped to the range $[k_{\min}, k_{\max}]$. The local matrix becomes

$$S_v(i,j) = \frac{W_v(i,j)\,\mathbb{1}[j \in N_i^v]}{\sum_{l \in N_i^v} W_v(i,l)},$$

and the SNF update $P_A^{(t+1)} = S_A P_B^{(t)} S_A^{\top}$ is otherwise unchanged. A small, tight cluster thus gets a small neighborhood instead of being padded with outsiders, while dense majority regions keep large neighborhoods and their noise suppression. The mutual-agreement condition stops hub majority patients from absorbing rare patients' edges. All three parameters ($\tau$, $k_{\min}$, $k_{\max}$) are tuned on validation macro-F1, with the same budget given to standard SNF's $k$.

The third component asks whether the graph fix and a loss fix are complementary or redundant. We will cross two graph options, standard SNF and adaptive SNF, with three training objectives: cross-entropy, cross-entropy plus soft-F1 with oversampling following Scholz et al., and DPGNN-style prototype training. The classifier is a shallow two-layer GCN. We choose a shallow network because Shwartz-Ziv et al. found depth hurts rare classes under imbalance. We will also include DeepMoIC's deep residual GCN to test that finding on graphs. Existing SNF tooling such as metasnf serves as the reference implementation that our modification must reproduce when $\tau = 0$ and the neighborhood size is fixed.

Baselines

The naive reference point is standard SNF with fixed $k$ (the DeepMoIC setting), a two-layer GCN, and plain cross-entropy, alongside a trivial classifier that always predicts the majority class. This reference shows how much of any reported accuracy comes from majority classes. It is run on identical stratified splits, and success means the proposal improves its rarest-class recall and macro-F1 by more than the across-seed standard deviation.

The strongest existing method on this question is MOGAT, a graph attention network over patient similarity graphs. On TCGA-BRCA with three data types it reports macro-F1 0.804, against 0.732 for SUPREME and 0.550 for MOGONET (MOGAT). It is the strongest competitor because its attention can in principle down-weight majority neighbors, which is exactly the effect our graph fix claims to need. We will re-run MOGAT on our splits with its published hyperparameters plus a tuning budget equal to ours, and quote 0.804 only as the published reference, since split protocols may differ. Beating it means higher macro-F1 and higher recall on the two rarest BRCA subtypes on matched splits, and further gains when MOGAT is run on our adaptive graph. A gain from swapping in our graph would show the graph is the bottleneck even for attention.

Further baselines are DeepMoIC and MOGONET, both run with their published architectures. For DeepMoIC the reported Pan-cancer figure is 84.28% accuracy; we found no reported balanced-accuracy or macro-F1 result on Lacuna for DeepMoIC or for MOGONET on BRCA beyond MOGAT's 0.550 figure. We also run DPGNN on the standard SNF graph, which is the strongest classifier-side remedy for graph imbalance, and XGBoost with balanced class weights on concatenated features. XGBoost uses the same feature preprocessing and receives an equal hyperparameter search.

Two ablations isolate the key ingredient. The first is standard SNF with $k$ swept down to $k_{\min}$ globally, which tests whether adaptivity matters or merely a smaller neighborhood does. The second is our adaptive neighborhoods combined with linear mean fusion (Deng et al.) instead of cross-diffusion, which tests whether cross-diffusion still earns its place once neighborhoods are fixed.

Why not just…?

Why not just concatenate features and use class-weighted XGBoost? This is a real competitor on small cohorts. MOGONET reports 0.815 accuracy on ROSMAP, beating XGBoost, but only in accuracy. In the clinical study of Wang et al., the authors merged T3 and T4 stages rather than confront the imbalance, so there is no evidence on how class-weighted boosting handles rare classes in multi-omics. We therefore run it as a full baseline. If it matches our rarest-class recall, the graph pipeline is not justified for this purpose, and we will report that.

Why not just add an imbalance-aware loss to the existing SNF-GCN pipeline? This is the most serious threat. Soft-F1 with oversampling raised balanced accuracy by up to 12.7% on glioma MRI (Scholz et al.). The two-by-three crossing of graph and objective settles the question. If soft-F1 or prototype training on the standard SNF graph matches adaptive SNF with cross-entropy, and the two do not add up, the graph-level story is wrong. The $h_c$ diagnostic should predict which outcome occurs: a loss cannot restore within-class edges that the graph never contains.

Why not just use a smaller global $k$? A small enough $k$ would stop forcing rare patients onto majority neighbors. It would also fragment large classes and weaken the cross-type agreement that suppresses noise, which is what made SNF robust in Duan et al.. The $k$-sweep ablation measures this tradeoff directly. If some single $k$ matches the adaptive scheme across all cohorts, we drop adaptivity and recommend tuning $k$ against macro-F1.

Why not oversample the rare class with SMOTE before building the graph? Synthetic points placed between real rare patients create artificial tight clusters that inflate within-class similarity without adding information. DPGNN reports that GraphSMOTE degrades steeply as skew grows (Wang et al.). We include SMOTE before graph construction as an extra arm under the same tuning budget.

Proposed Experiments

The primary cohorts are TCGA-BRCA, METABRIC, and TCGA-KIPAN, each using five stratified splits and three seeds. We report macro-F1, balanced accuracy, and per-class recall, with overall accuracy given only for comparison to prior papers. TCGA Pan-cancer, with 28 classes of very uneven size, tests whether the effect grows with the number of rare classes. ROSMAP serves as a near-balanced control on which the fix should change little.

Three checks decide the outcome. The first is whether rare-class $h_c$ falls under standard SNF once class size drops below $k$ in the controlled subsampling, and whether the adaptive neighborhood prevents that fall. The second is whether gains in $h_c$ predict gains in rare-class recall across cohorts and graph variants. The third is whether adaptive SNF improves rare-class recall over the best of soft-F1, DPGNN, MOGAT, and the best global $k$, and still adds to them when combined. If the first check fails, the premise is false. If the third fails while the first holds, the conclusion is that fused graphs do lose rare subtypes but downstream remedies already compensate. Either result corrects how the field evaluates multi-omics graph classifiers, which currently report only overall accuracy.