Predicting which microRNAs, long non-coding RNAs or microbes are linked to which diseases is usually posed as filling in missing entries of a sparse association matrix. Most methods make one assumption about these graphs: connected or similar things share disease links, which biologists call "guilt by association". Graph neural networks for this task inherit that assumption through smoothing, meaning each node repeatedly averages its neighbours' features. This note argues that in these graphs, over-smoothing (the collapse of all node representations toward one shared vector as layers are stacked) has a concrete and checkable form. It turns predictions into a popularity ranking, where the score of a pair is roughly the product of the two nodes' degrees, their number of known associations. That ranking looks good on average AUC and is weakest on the rare diseases and newly characterised RNAs where predictions matter most.
The proposal is a small, analysable model. It learns a signed filter over the singular-value spectrum of the degree-normalised association matrix. Separately, for each similarity network, it learns how much to smooth versus sharpen, guided by a measured, label-free homophily score. The model contains path-counting methods like KATZ as a special case. Its central test is narrow: does allowing negative (sharpening) filter coefficients improve prediction for low-degree entities while keeping the representation from collapsing toward rank one?
Core Claim
Dual-pass graph filtering means running a smoothing (low-pass) channel and a sharpening (high-pass) channel side by side. Methods that do this, such as FB-GNN, PC-Conv, RASH and H2SGNN, were built for node classification. They do not transfer directly to association prediction, for a simple spectral reason.
On a purely bipartite block, such as miRNA–disease links, the eigenvalues of the normalised adjacency come in pairs $\pm\sigma$. Here $\sigma$ ranges over the singular values of the normalised association matrix $\tilde B = D_m^{-1/2} B D_d^{-1/2}$, where $B$ is the binary training association matrix and $D_m, D_d$ are the diagonal degree matrices. The "smoothest" and "sharpest" Laplacian components therefore share the same singular pair and differ only by a sign flip on one side. A cross-type link score sees only the odd part of any filter. The low-pass/high-pass split collapses, and the choice that actually matters is which singular values to weight. We therefore score pairs as
$$ \hat Y = \sum_{j=0}^{J} \theta_j\, \tilde B(\tilde B^\top \tilde B)^j = U\, g(\Sigma)\, V^\top, \qquad g(\sigma)=\sum_{j=0}^{J}\theta_j\,\sigma^{2j+1}. $$
Here $U\Sigma V^\top$ is the singular value decomposition of $\tilde B$, and the coefficients $\theta_j$ are learned and may be negative. The term $\tilde B(\tilde B^\top \tilde B)^j$ counts degree-normalised walks of length $2j+1$, so positive geometrically decaying $\theta_j$ recovers a KATZ-style walk count.
Over-smoothing is now explicit. As walk length grows, $g$ is dominated by the top singular value $\sigma_1 = 1$. For a connected graph, the matching singular vectors are proportional to the square roots of the degrees. Deep smoothing therefore converges to $\sqrt{d_m d_d}$, a pure popularity ranking. This is rank-one collapse in the sense of Are we measuring oversmoothing correctly?, where Effective Rank tracked accuracy with average Pearson correlation 0.91. That paper motivates our choice of diagnostic. The collapse is also the bipartite form of the "pseudo-heterophily" that Two Sides of the Same Coin identifies: nodes with low relative degree drift first. That analysis predicts where gains should appear.
The similarity networks break bipartiteness, and there a genuine smooth-versus-sharpen choice exists. These include miRNA sequence or functional similarity and disease semantic similarity from the Medical Subject Headings (MeSH) vocabulary of disease terms. For each similarity block $S$ we compute an association homophily score. It is the correlation between the entries of $S$ and the overlap (cosine) of the corresponding training association rows. A positive score means the similarity really does predict shared diseases. A score near zero or negative means guilt by association fails for that data source. Each block then receives an FB-GNN-style mix of smoothing and sharpening. The mixing weights start from that score and are learned from there.
Research Program
The work has three parts.
The first is an audit, done before any model is built. On each benchmark, we project held-out associations onto the singular directions of $\tilde B$ and measure how their mass spreads across singular values. We do this separately for high-, middle- and low-degree diseases. We also compute association homophily for every similarity source. The audit answers whether held-out links for rare entities sit in the mid-spectrum, where low-pass methods give them little weight. It also answers whether any commonly used similarity is weakly or negatively homophilous.
One known confound must be handled first. The Gaussian Interaction Profile (GIP) kernel, a similarity computed from each entity's association row and used by PBMDA, KATZNCP and HG-LGBM, is built from the associations themselves. We recompute it from training folds only, so it cannot leak test links.
The second part is the model. It combines the signed association filter above with per-similarity-block dual-pass mixing and a low-order polynomial on the full augmented operator. The parameter count is in the dozens to hundreds, it trains with a logistic loss on sampled pairs, and it runs in seconds on graphs of this size. Keeping the model linear in the graph operator is deliberate. Every learned $g(\sigma)$ can be plotted, and every gain can be attributed to a spectral band.
The third part is diagnosis. Throughout training and across polynomial order $J$, we track three quantities: the Effective Rank of $\hat Y$, the Spearman correlation between $\hat Y$ and the degree-product ranking, and AUPR within degree buckets. The claim is falsified if negative coefficients improve AUC but not low-degree AUPR, or if the correlation with popularity does not fall.
Baselines
The naive reference point is degree-product (preferential attachment) scoring, together with random scoring. Under random-negative evaluation, degree-product may already reach high AUC. If so, that alone recalibrates how to read the published 0.93–0.98 figures.
The strongest classical competitor is KATZ walk counting as in KATZNCP, which reports leave-one-out cross-validation (LOOCV) AUC 0.9325 on HMDD v2.0. A second is length-3 path weighting as in PBMDA, which reports AUC 0.9172. These are the most direct threat, because our model nests them.
The strongest neural competitors are each benchmark's published model. On the lncRNA tripartite network this is GCNLDA, which reports AUC 0.959, AUPR 0.2233 and 91.5% recall in the top 30. On CircR2Disease it is MSMCDA, with AUC 0.976. On HMDAD it is HG-LGBM, with AUC 0.9757 ± 0.0032. We also include closed-form Personalized PageRank diffusion in the style of TAGDN, and FB-GNN and H2SGNN ported to link prediction with an inner-product decoder.
Published numbers come from each paper's own splits and negative sampling, so they are not directly comparable. We will rerun every baseline we can reimplement under one shared protocol. All models get the same training-only similarities, the same folds and the same hyperparameter search budget, set at fifty trials per method. The published figures serve only as sanity anchors for each reimplementation.
The key ablation constrains all $\theta_j \ge 0$. This gives a learned smoothing-only filter, a tuned KATZ. A second ablation fixes the similarity mixing to pure smoothing. A third drops the homophily-based initialisation.
Beating the baselines means two things. The first is a statistically reliable gain in AUPR on the lowest-degree tercile and under disease-held-out evaluation, with no loss in overall AUC. The second is a lower correlation with popularity than the nonnegative ablation. Overall AUC parity combined with a low-degree gain would still count as a positive result.
Why not just…?
Why not just tune KATZ or a path method? This is the most serious alternative, because a well-tuned walk count may capture most of the signal. The nonnegative ablation is exactly a learned KATZ. If it matches the signed model on low-degree AUPR and cold-start recall, then the dual-pass ingredient adds nothing in this domain, and we will report that as the finding.
Why not just run gradient-boosted trees on similarity profiles? Boosted trees can learn non-linear rules without any graph propagation, so they cannot over-smooth at all. HG-LGBM and DMAGCL report that standalone tree and kernel classifiers trail graph-based representations, but those comparisons used each paper's own features. We will run LightGBM on concatenated training-only similarity rows under our protocol. We will also check whether its errors share the popularity bias, since profile rows of rare entities are nearly empty.
Why not just use a shallow two-layer graph network? Shallow depth limits smoothing. However, FB-GNN shows that even two-layer aggregation discards useful high-frequency signal when homophily is low. Our degree-product correlation diagnostic measures directly whether a shallow network trained for this task already ranks by popularity.
Why not just port H2SGNN or RASH? Both learn separate smoothing and sharpening filters on heterogeneous graphs. The bipartite argument above predicts that, applied to the association block, their two channels are redundant up to sign. Running them as baselines tests that prediction. If they match our model, the singular-value view is still the explanation, but it is no longer a reason to use our method.
Proposed Experiments
The datasets are HMDD v2.0 (495 miRNAs, 383 diseases, 5,430 associations), the HMDD v3.0 release (HMDD v3.0, 32,281 entries), the lncRNA–disease–miRNA network used by GCNLDA, CircR2Disease, and HMDAD. Each runs under three protocols:
- five-fold random holdout;
- disease-held-out cold start, mirroring PBMDA's glioblastoma simulation;
- temporal validation, training on HMDD v2.0 associations and testing on links that are new in v3.0.
Metrics are AUC, AUPR, recall in the top 30, AUPR within degree terciles, the Effective Rank of $\hat Y$, and the Spearman correlation with degree-product ranking.
Three checks decide the outcome. The first: do the learned $g(\sigma)$ curves assign negative or mid-spectrum weight, and does that pattern repeat across datasets? The second: does the signed model beat the nonnegative ablation on low-degree and temporal AUPR? The third: across all models and depths, does the correlation with degree-product predict low-degree AUPR better than Effective Rank alone? If yes, the popularity-collapse account of over-smoothing is supported. If not, the account fails, even if the model wins on average AUC.
References
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks
- PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering
- Enhancing Homophily-Heterophily Separation: Relation-Aware Learning in Heterogeneous Graphs
- Addressing Heterogeneity and Heterophily in Graphs: A Heterogeneous Heterophilic Spectral Graph Neural Network
- Are we measuring oversmoothing in graph neural networks correctly?
- Two Sides of the Same Coin: Heterophily and Oversmoothing in Graph Convolutional Neural Networks
- PBMDA: A novel and effective path-based computational model for miRNA-disease association prediction
- KATZNCP: a miRNA-disease association prediction model integrating KATZ algorithm and network consistency projection
- HG-LGBM: A Hybrid Model for Microbiome-Disease Prediction Based on Heterogeneous Networks and Gradient Boosting
- Graph Convolutional Network and Convolutional Neural Network Based Method for Predicting lncRNA-Disease Associations
- Predicting circRNA-disease associations with shared units and multi-channel attention mechanisms
- TAGDN: type-aligning graph diffusion network for capturing long-range dependencies in heterogeneous graphs without over-squashing and over-smoothing
- DMAGCL: A dual-masked adaptive graph contrastive learning framework for predicting circRNA-drug sensitivity
- HMDD v3.0: a database for experimentally supported human microRNA–disease associations