Antibodies and designed binders recognize patches on a pathogen's surface proteins called epitopes. The epitopes worth targeting are the conserved ones — fusion stems, catalytic pockets, genome-packaging interfaces — because the pathogen cannot mutate them without losing function, as documented for influenza HA stem regions, coronavirus protease active sites, and nucleocapsid RNA-binding domains (Chen et al., Cao et al., Fan et al.). Yet computational predictors reliably miss them. Epitope prediction on the standard benchmark reaches only an area under the precision–recall curve of 0.277 (Del Vecchio et al.), and the failures are not random: predictors favor prominent, densely packed, variable surface loops and miss recessed, sparsely connected conserved pockets.
I propose that this failure is a known disease with a known cure. Epitope predictors are message-passing networks over graphs whose nodes are amino-acid residues (Del Vecchio et al., Kong et al.), and graph learning theory has proved, in two separate lines, that message passing (a process where each node updates its representation by averaging its neighbors') systematically favors densely connected nodes and nodes close to training examples (Zhang et al., Ma et al.). Conserved epitopes are disadvantaged on exactly both axes: they sit in sparse, recessed neighborhoods, and they are structurally far from the immunodominant regions that dominate SAbDab-style training data. No existing predictor corrects for either. The plan is to port two proven structural-bias mitigations — degree-inverted aggregation and counterfactual mutation invariance — into the leading epitope architecture, and to test whether conserved-epitope recall rises without sacrificing overall precision.
Core Claim
The systematic under-detection of conserved epitopes is a structural-bias phenomenon, not a data-volume or class-imbalance phenomenon, and it can be mitigated by architectural changes imported from graph fairness research. Three facts from the literature motivate this. First, Zhang et al. proved that a region's embedding quality under graph learning is governed by its local Laplacian spectrum: sparse regions converge slowly into blurry embeddings, and their fix — coarsening with edge weights proportional to $1/(d_u + d_v)$, where $d_u$ is the graph degree of node $u$ — closed the accuracy gap across classes. Second, Ma et al. proved test error rises monotonically with a node's distance from training labels, and Su et al. showed high structural awareness deepens this "accuracy valley" around training hubs. Third, Cheng et al. showed that spurious signal leaks across edges into protected nodes and can be severed with counterfactual graphs that resample the spurious attributes. Translated to immunology: dense protruding loops converge to crisp, well-classified embeddings while recessed pockets blur, and the mutational noise of variable surface residues leaks into the embeddings of adjacent conserved residues. The proposal builds a predictor, EpiParity, that addresses both, and predicts — falsifiably — that the recall gain concentrates on conserved, low-degree, training-distant residues.
Research Program
EpiParity keeps the backbone of Neural message passing for joint paratope-epitope prediction (its bipartite cross-attention design is the strongest published epitope architecture) and changes only how the antigen graph is aggregated and trained. There are two ingredients, plus one diagnostic.
The first ingredient is degree-inverted aggregation. In the antigen residue graph (residues connected if within roughly 10 Å), protruding variable loops have high contact degree while recessed conserved pockets have low degree. Following the coarsening analysis of Zhang et al., each message-passing step reweights edges as
$$w(u,v) \propto \frac{1}{d_u + d_v}$$
so that messages entering sparse pockets are concentrated rather than diluted by averaging over many neighbors. This directly transplants the mechanism that reduced class-accuracy gaps on citation and e-commerce graphs (Zhang et al.) onto residue graphs, where the disadvantaged "communities" are the recessed conserved sites.
The second ingredient is mutation-counterfactual invariance training. Cheng et al. built counterfactual graphs by resampling spurious attributes and severing the leakage paths; their construction scales quadratically via a KNN rebuild, which they flag as the main scalability limitation. We avoid that cost entirely: an antigen's variable residues (identified by alignment entropy across homologs, exploiting the fitness-constrained trajectories of Hu et al.) are resampled in place from homolog sequences, leaving conserved residues and the graph topology untouched. The predictor is then penalized whenever its representation of a conserved residue shifts under this counterfactual:
$$\mathcal{L}_{\text{inv}} = \sum_{r \in \mathcal{C}} \lVert h_r(G) - h_r(\tilde{G}) \rVert^2$$
where $h_r$ is the embedding of residue $r$, $\mathcal{C}$ is the conserved set, and $\tilde{G}$ is a mutation-counterfactual antigen. This severs the leakage of strain-specific surface signal into conserved-site embeddings, at $O(N)$ cost per counterfactual rather than the $O(N^2)$ of Cheng et al..
The third component is a diagnostic, not a mechanism. Following Ma et al., who showed accuracy descending as a staircase with distance from training labels, we will measure whether existing predictors show the same staircase on antigen graphs — recall of true epitope residues bucketed by hop distance to the nearest training epitope residue, and separately by residue degree. If the staircase is flat in existing models, the premise of the proposal is wrong and we will report that; the theory of Ma et al. and Su et al. predicts it will not be. A light mitigation rides on the diagnostic: distance-stratified sampling of training complexes, since Ma et al. showed that sampling training labels centrally biases predictions toward dominant classes.
This proposal is deliberately distinct from the nearest existing bridges. Geometric epitope and paratope prediction handles rotational invariance and coordinate noise, and even observes that its graph model prefers sparser antigen regions — but does not measure or mitigate degree bias, and requires both bound partners as input. Del Vecchio et al. address label imbalance via cross-attention, not topology. CHIMERA-Bench proves that models collapse on held-out epitope groups and novel antigen folds but changes no architecture; it supplies our evaluation regime, not our method. The APL framework of Structural Framework for Analysis of CD4+ T-Cell Epitope Dominance in Viral Fusion Proteins supplies the biophysical insight that conserved segments must also survive proteolysis and remain solvent-accessible — a constraint we use only to define the conserved-epitope test strata.
Proposed Experiments
The premise check comes first: run the staircase diagnostic on the published Epi-EPMP model over its own epitope benchmark (132 training/validation and 30 test complexes from SAbDab and Docking Benchmark v5, residues featurized exactly as in Del Vecchio et al.), bucketing test residues by hop distance to training epitope residues and by residue degree. The decisive figure is a recall-versus-distance curve for the baseline versus EpiParity: the claim predicts a steep staircase for the baseline and a flattened one for ours.
The main comparison runs on that benchmark and on the GEP test set of 49 SAbDab complexes with at most 70% sequence identity to training (Geometric epitope and paratope prediction), reporting AUC PR, AUC ROC, and MCC both overall and stratified by conservation level and residue degree. The headline metric is conserved-epitope recall — recall restricted to epitope residues that are conserved across homologs and recessed (low degree or low solvent accessibility) — because that is the quantity that matters for variant-resilient targeting. Overall precision must not degrade materially; a predictor that shouts "epitope" everywhere would win recall trivially, so precision at matched recall is the honest gate.
The external validity test uses a curated conserved-target panel assembled from the structural biology in this file: influenza HA stem and hydrophobic fusion machinery (Chen et al., Weng et al.), coronavirus M$^{pro}$ and PL$^{pro}$ active sites (Yang et al., Amporndanai et al.), the nucleocapsid RNA-binding interface (Fan et al.), and TNF$\alpha$ hydrophobic pockets (Weng et al.). No reported result found on Lacuna for any existing epitope predictor on this panel, so we generate all baseline numbers ourselves under identical conditions.
Finally, downstream design: condition the MEAN-style equivariant generator (Kong et al.) on EpiParity-predicted conserved epitopes and evaluate on the epitope-group and antigen-fold splits of CHIMERA-Bench, reporting epitope F1, $F_{nat}$, and DockQ. As reference points only: MEAN reports 58.29% amino-acid recovery on CDR-H1 with 0.98 Å RMSD on its own SAbDab benchmark (Kong et al.), but no reported result found on Lacuna for per-split CHIMERA-Bench values of any model, so the comparison is internally generated. The downstream hypothesis is that part of the held-out-epitope degradation CHIMERA-Bench documents is distance-and-degree bias, and that debiased conditioning recovers some of it.
Baselines
The naive reference point is a conservation-only scorer: rank residues by inverse alignment entropy across homologs, with no learned structure. This is the "do nothing" floor, and the APL analysis of Structural Framework for Analysis of CD4+ T-Cell Epitope Dominance in Viral Fusion Proteins — where conserved segments that are buried or proteolytically destroyed never become epitopes — predicts it will have respectable apparent recall but poor precision on accessible conserved sites.
The strongest existing competitor is Epi-EPMP from Del Vecchio et al., which reports an epitope AUC PR of 0.277 on its benchmark, versus 0.212 for PECAN. It is the strongest because it is the best published number on exactly our task and because it already targets the class-imbalance problem via bipartite cross-attention — meaning any gain we show cannot be attributed to imbalance handling. We will re-run it from the authors' protocol on the same splits, the same 62-dimensional residue features, the same graph construction (15 nearest neighbors within 10 Å), and the same hyperparameter search budget as EpiParity, differing only in the aggregation weights, the invariance loss, and the stratified sampling. Beating it means strictly higher conserved-epitope recall at matched or better overall precision, with overall AUC PR not below its reported 0.277 by more than noise.
A second competitive baseline is the outer-structure surface model O-GEP with Heat Kernel Signatures from Geometric epitope and paratope prediction, which showed MCC gains of 0.09 and 0.12 over its non-HKS variants; it represents the point-cloud route and, notably, was observed to localize predictions on continuous surface patches, which could plausibly favor conserved stems.
Three ablations isolate the ingredients. The first replaces degree-inverted aggregation with standard uniform aggregation while keeping everything else, isolating ingredient one. The second keeps uniform aggregation but adds the mutation-counterfactual invariance loss, isolating ingredient two. The third — and the one a skeptic will demand — is the backbone plus focal loss with class weights, which tests whether any observed gain is just better loss weighting rather than structural debiasing; Zhang et al. proved disparity is rooted in the Laplacian spectrum rather than label counts, and PECAN's weighted-loss result of 0.212 AUC PR (Del Vecchio et al.) supports this, but the ablation settles it empirically on our task.
Why not just…?
Why not just use focal loss or class weighting? This is the most serious alternative, because the failure mode — near-zero recall on a small residue class — does look like textbook imbalance. The evidence against it is Zhang et al., who proved that performance disparity across graph regions is driven by the local Laplacian spectrum and edge density, which loss weights do not touch, and the empirical datum that PECAN, a weighted-loss method, reached only 0.212 AUC PR for epitopes versus 0.277 for the architecturally adapted EPMP (Del Vecchio et al.). Still, the focal-loss ablation is in the plan precisely because if it matches EpiParity on conserved-epitope recall at matched precision, the architectural claim is dead and the simpler method wins.
Why not just score sequence conservation and skip graphs? Conservation is computable from an alignment in minutes. But the APL analysis of Structural Framework for Analysis of CD4+ T-Cell Epitope Dominance in Viral Fusion Proteins shows conservation alone fails as a targeting signal — presentation is dictated by 3D stability, B-factors, and conformational state, with its own framework reaching positive predictive values up to 48% for HIV Env only after adding structure — and CHIMERA-Bench shows sequence-level quality does not translate into interface quality ($F_{nat}$, DockQ). The MEAN ablations reinforce this: removing structural antigen context made the binder model fail entirely at affinity optimization (Kong et al.). The conservation-only baseline is included anyway, as the honest floor.
Why not just drop the graph and use point clouds or plain MLPs? If message passing causes the bias, removing it seems natural. But Geometric epitope and paratope prediction found its residue-graph model captures inner interactions that the point-cloud model misses, and Kong et al. showed that representations using only pairwise distances, without directional coordinate tracking, distort backbone geometry (~3.2 Å RMSD versus 0.98 Å for the equivariant graph model). Meanwhile, Zhang et al. note that an MLP shows a smaller class spread than graph methods but at substantially lower absolute accuracy — trading the disease for weakness. The genuine threat inside this alternative is the O-GEP surface model with Heat Kernel Signatures, which is why it is a full baseline rather than a footnote; if it matches EpiParity on the conserved-target panel, the residue-graph framing of the bias is the wrong one, and that would itself be a finding worth reporting.
Decisive Checks
The proposal stands or falls on three numbers. First, the staircase diagnostic: a monotone recall drop with distance-to-training-epitopes in Epi-EPMP, flattening under EpiParity — if the baseline staircase is flat, the whole premise is refuted. Second, conserved-epitope recall on the main benchmark at matched precision, against the 0.277 AUC PR bar of Del Vecchio et al. and the focal-loss ablation. Third, transfer to the conserved-target panel of HA stem, protease active sites, and nucleocapsid interfaces — the settings where variant-resilient targeting actually matters. A secondary check extends the evaluation to conformational ensembles, since Structural Framework for Analysis of CD4+ T-Cell Epitope Dominance showed neutral-pH and low-pH HA conformations differ predictively: the counterfactual-invariance loss should, as a side effect, make predictions stable across pre- and post-fusion states, a robustness test no current rigid-pose model passes.
References
- Asprellcosides B of Ilex asprella Inhibits Influenza A Virus Infection by Blocking the Hemagglutinin- Mediated Membrane Fusion
- High-throughput screening of dual-target inhibitors for SARS-CoV-2 main protease and papain-like protease from Chebulae Fructus: in silico prediction and experimental verification
- The highly conserved RNA-binding specificity of nucleocapsid protein facilitates the identification of drugs with broad anti-coronavirus activity
- Neural message passing for joint paratope-epitope prediction
- Conditional Antibody Design as 3D Equivariant Graph Translation
- Understanding Community Bias Amplification in Graph Representation Learning
- Subgroup Generalization and Fairness of Graph Neural Networks
- On the Topology Awareness and Generalization Performance of Graph Neural Networks
- Causality-Inspired Disentanglement for Fair Graph Neural Networks
- Dissecting the Dynamics of HIV-1 Protein Sequence Diversity
- Geometric epitope and paratope prediction
- CHIMERA-Bench: A Benchmark Dataset for Epitope-Specific Antibody Design
- Structural Framework for Analysis of CD4+ T-Cell Epitope Dominance in Viral Fusion Proteins
- Design of minibinder proteins as universal antagonists against canine and human TNFα
- Structure-assisted design of inhibitors targeting coronavirus main proteases
- Inhibition mechanism of SARS-CoV-2 main protease by ebselen and its derivatives