Many human proteins live in more than one place in the cell. One source estimates about 20% (DeePSLoc). Predicting all of a protein's compartments from its folded 3D shape is therefore a multi-label problem. A common way to feed shape to a model is a backbone graph, in which each amino-acid residue is a node and edges join chain neighbors and residues that sit close in space. A graph neural network (a model that updates each node by averaging in its neighbors) then summarizes the graph. The difficulty is that the clues to location are often small and local: a short N-terminal targeting peptide, a basic nuclear signal, one transmembrane helix. Each round of neighbor averaging blurs such clues. The final step, averaging all residues into one vector, keeps only the protein-wide mean, which is the smoothest summary possible. This is the "over-smoothing" problem in graph learning, here applied to protein graphs. Compartments that depend on local clues should suffer most, and those compartments are also the rare ones.
We propose to measure this blurring directly on protein backbone graphs and then fix it at the readout. The fix gives each compartment its own small attention module. That module chooses both which residues to read and at which spatial scale to read them, picking among smooth, intermediate and sharp versions of the residue features. These versions are computed with fixed, parameter-free graph filters rather than by stacking more layers. The claim is narrow and testable. On the CAPSUL benchmark, compartments whose clues are local should gain the most from access to high-frequency (sharp, local-contrast) features. The gain should be largest for rare compartments. It should appear only when the choice of scale is made separately for each label.
Core Claim
Structure-based localization models evaluated on CAPSUL are shallow, and CAPSUL never asks why (CAPSUL). We hypothesize two things. First, depth hurts these models because residue features lose dimensionality. This collapse falls hardest on the protein's "extremities": low-contact termini and disordered tails, which is where many targeting peptides sit. Second, a single shared pooled vector forces every compartment to read the same smoothed summary.
Our readout replaces that shared vector. Let $h_v$ be the final-layer feature of residue $v$ and $\tilde{A}$ the normalized backbone adjacency. We build a small bank of filtered features: a low-pass channel $\tilde{A}^K h$; diffusion-wavelet band-pass channels $(\tilde{A}^{2^{b-1}} - \tilde{A}^{2^b})h$, following Geometric Scattering Attention Networks; and the high-pass "diversification" channel $(I-\tilde{A})h$, following FB-GNN. Call these filters $F_b$. Each compartment $c$ gets its own representation
$$z_c = \sum_{b}\sum_{v} \alpha_{c,b,v}\,(F_b h)_v, \qquad \alpha_{c,\cdot,\cdot} = \mathrm{softmax}_{b,v}\big(q_c^\top W_b (F_b h)_v\big),$$
where $q_c$ is a learned query for compartment $c$ and $W_b$ is a projection for channel $b$. A per-label logistic output then reads $z_c$. The attention weights also give an interpretable map: for each label, they show where on the structure the decision came from and at which scale.
What prior work supplies, and what it leaves open
Three papers set the application. CAPSUL motivates the work. It shows that 3D geometry carries the signal, because randomizing coordinates collapses structural models. It also shows that structure helps minority compartments. But it never studies depth, rank or readout in its GNN backbones (CAPSUL). GearNet-Edge supplies the encoder: a relational residue graph with sequential, radius and nearest-neighbor edges, plus edge-to-edge messages that encode angles. It was kept to a few layers and evaluated on function, not localization (GearNet). DeepLoc 2.0 is the strongest sequence-only approach: protein language model embeddings with attention pooling over residues (DeepLoc 2.0). It constrains the proposal, because any gain we report must survive comparison with a well-tuned version of that pooling.
The filtering ideas come from single-label node classification. FB-GNN adds a learnable high-pass channel to a GCN, a standard graph convolutional network. GSAN attends over wavelet channels for each node. Both were tested on citation and web graphs and never on graph-level, multi-label targets. Our use differs in two ways. The filters act at the readout of a whole protein graph rather than inside every layer. And the selection over channels is made per label rather than per node. hpGAT's strategy of capturing multi-hop structure without deep stacking (hpGAT) supports a design choice here: because our wavelets are fixed powers of $\tilde{A}$, we avoid scaling up weights. Backward Oversmoothing shows that scaling weights up to escape forward collapse destabilizes gradients instead.
Our diagnostics come from Are we measuring oversmoothing correctly?. That paper shows Dirichlet energy, a standard smoothness measure, can stay flat while accuracy drops in trained GNNs. Effective rank (the entropy-based count of significant singular values of the feature matrix) instead tracks accuracy, with an average Pearson correlation of 0.91. That evidence comes from node classification, so whether rank tracks per-label accuracy on protein graphs is itself one of our questions.
The skeleton literature supplies the lens for non-uniform collapse. In human-pose graphs, rigid hubs and loose extremities behave differently; ST-GCN partitions neighbors relative to the body's center for this reason. We carry the same distinction to proteins: a dense hydrophobic core versus low-contact termini.
Research Program
The first stage is diagnostic. We train GearNet-Edge on CAPSUL at depths 2, 4, 6, 8, 12 and 16, with and without residual connections. At each layer we record effective rank and Dirichlet energy. Rank is measured three ways: over all residues, separately over core residues versus extremity residues (defined by contact degree in the radius graph and by AlphaFold confidence), and across proteins on the pooled vectors. Alongside rank we record per-label test performance. We also annotate known targeting regions, using UniProt signal and transit-peptide features where CAPSUL's UniProt-derived entries carry them. On these regions we fit linear probes on residue features to ask whether "this residue lies in a targeting peptide" stays decodable as depth grows.
The prediction is specific. Extremity rank and probe accuracy should fall earlier than core rank. Compartments tied to N-terminal or helix signals, such as mitochondria, secretory-pathway compartments and the Golgi apparatus (which CAPSUL links to transmembrane helices), should degrade first.
The second stage is the method: the label-wise frequency readout placed on top of a shallow GearNet-Edge. We test node features of two kinds, structural features alone and ESM-2 per-residue embeddings. The ESM-2 variant exists to separate what structure adds from what the readout adds.
The third stage is interpretation. We check whether the attention for each label concentrates on annotated signal regions and on the high-pass channel for local-signal compartments, and on low-pass channels for compartments thought to depend on global properties.
Baselines
The naive reference point is GearNet-Edge at its default shallow depth with mean pooling and independent per-label sigmoid outputs, run as CAPSUL configures it. Every gain is reported relative to it.
The strongest existing structural competitor is the best CAPSUL backbone in that paper's own comparison, among CDConv, GearNet-Edge, Graph Mamba and Graph Transformer, trained with CAPSUL's class reweighting. No reported result found on Lacuna for CAPSUL per-class or aggregate scores, so we rerun it.
The strongest sequence competitor is DeepLoc 2.0-style attention pooling over ESM-2 residue embeddings. Its reported sorting-signal accuracy, 79% across nine signal types on its own benchmark (DeepLoc 2.0), is not comparable to CAPSUL compartments. We retrain it on CAPSUL splits rather than quote it.
Fairness rules apply to all models. Every model sees the same splits, the same AlphaFold structures and the same loss, and gets the same tuning budget: a fixed number of hyperparameter trials per model and matched parameter counts within 10%. Results are averaged over five seeds.
Two ablations isolate the key ingredient. In the first, the same readout uses a single attention module shared across labels; this tests whether per-label selection matters. In the second, the readout per label is restricted to the low-pass channel only; this tests whether high frequencies matter. A third ablation removes residue attention and keeps channel attention; this tests whether "where" or "which scale" carries the gain.
Beating the field means three things. Macro-averaged AUPRC across the 20 compartments must exceed the best rerun baseline by more than the seed standard deviation. The improvement must be concentrated in the rarest third of compartments. And both ablations must fall back toward baseline.
Why not just…?
Why not just use a protein language model with attention pooling? This is a real threat. Per-residue attention already lets local signals bypass averaging, and language models encode a great deal of structure. CAPSUL's coordinate-randomization result and its minority-class findings argue that geometry adds information (CAPSUL). Even so, a per-label attention pool over ESM-2 features, with no graph at all, is in our baseline set. If it matches the full method, we will report that the backbone graph is not needed for this gain. The ESM-2-on-graph variant then tells us whether the graph filters add anything beyond the embeddings.
Why not just keep the GNN shallow, or concatenate all layers? Shallow models avoid collapse but see only small neighborhoods. Concatenating every layer's output gives the classifier a crude multi-scale view, so it is a real competitor. We include it, with per-label attention over the concatenated features, as an explicit baseline. Our readout differs in two ways: it carries an explicit high-pass channel that concatenation lacks, and it selects scales per label. If concatenation ties us, the conclusion is that per-label readout matters and the filter design does not, and we will say so.
Why not just use a Graph Transformer, which attends globally and sidesteps neighbor averaging? CAPSUL already evaluates Graph Transformers and Graph Mamba for long-range context. The open question is whether they retain local clues for rare labels, so we report their per-label and per-region rank alongside ours. If a Graph Transformer already preserves extremity rank and wins on rare compartments, our diagnosis is wrong for this domain.
Why not just model label correlations with a label graph, or train independent classifiers? Label dependencies are real; nucleolus implies nucleus (kernel combinations). On HPA-18 images, the label-graph method ML-GCN reaches 60.3% subset accuracy against 61.2% for DeePSLoc (DeePSLoc), a gap too small to show that correlation modeling decides the outcome. We keep independent outputs to isolate the readout and add hierarchy-consistent post-processing as a secondary check.
Proposed Experiments and Decisive Checks
The primary benchmark is CAPSUL: 20,181 human proteins, 20 compartments, AlphaFold2 structures. The secondary benchmark is the DeepLoc SwissProt cross-validation set with its homology-partitioned independent HPA test set, using AlphaFold structures. Metrics are macro and per-class AUPRC, macro-F1 and subset accuracy, with effective rank and probe accuracy as diagnostics.
Three checks decide the outcome. The first is depth: if extremity effective rank does not fall faster than core rank, and per-label rank does not correlate with per-label performance across depths, the over-smoothing diagnosis fails for protein graphs. We would then present the readout only as a pooling improvement. The second is the per-label high-pass ablation. If restricting labels to the low-pass channel costs nothing on local-signal compartments, the frequency story is false. The third is the sequence control. If per-label attention pooling on ESM-2 alone matches the structural readout on minority compartments, the structural contribution is not supported.
A positive result requires all three checks to point the same way. It would be the first evidence that over-smoothing on biological backbone graphs is non-uniform and label-specific, together with a cheap fix that does not require deeper networks.
References
- Multi-marginal Contrastive Learning for Multilabel Subcellular Protein Localization
- CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization
- Geometric Scattering Attention Networks
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks
- Protein Representation Learning by Geometric Structure Pretraining
- DeepLoc 2.0: multi-label subcellular localization prediction using protein language models
- hpGAT: High-Order Proximity Informed Graph Attention Network
- Backward Oversmoothing: why is it hard to train deep Graph Neural Networks?
- Are we measuring oversmoothing in graph neural networks correctly?
- Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition
- Improved functional prediction of proteins by learning kernel combinations in multilabel settings