Function-Gated Consensus Prediction of Conserved Epitopes

When a virus mutates, most of its surface changes freely, but a few sites cannot: the catalytic pocket of a protease, the RNA-binding groove of a nucleocapsid, the fusion machinery of influenza hemagglutinin. These invariant sites are the "choke points" that broad-spectrum antivirals and universal vaccines want to hit, but computational pipelines rarely find them well. Single predictors of antibody-binding sites on protein surfaces either miss them or take hours per complex, and even when a predictor outputs a contact map, nothing in the pipeline tells us whether the targeted residues are doing anything the pathogen depends on. I propose to fix both problems at once with a single score: a per-residue consensus across orthogonal epitope predictors, multiplied by a functional gate that asks whether the residue sits in a region the pathogen's own biology annotates as essential.

The core mechanism is a deliberate transfer of two ideas from neighboring fields that have never been combined at the residue level. From ensemble target prediction, I borrow the CoBDock lesson that combining several imperfect predictors with orthogonal assumptions — as CoBDock does for small-molecule docking with four docking engines and two cavity detectors — beats every constituent tool; the same complementary-error logic was shown earlier by Sen et al. for binding-site prediction and by PARCE, where a majority vote across six scoring functions was the only way to keep peptide optimization monotonic. From functional enrichment analysis, I borrow the statistical machinery of overrepresentation — the hypergeometric test that asks whether a set of molecules is unusually concentrated in a curated functional category — but I apply it to residues, not genes: the "gene set" becomes a structural neighborhood on the antigen surface, and the annotation universe becomes residue-level functional terms (catalytic-activity residues, RNA-binding residues, pathway membership of the viral protein). The resulting method predicts which surface patches are simultaneously (a) antibody-accessible by consensus of several predictors, (b) evolutionarily conserved, and (c) functionally loaded — which is precisely the operational definition of a conserved choke-point epitope.

Core Claim

No existing pipeline computes all three of these signals jointly at the residue level. The closest works each stop one step short. WALLE on the AsEP benchmark predicts antibody-specific epitopes with a hybrid protein-language-model plus graph network and reaches an MCC of 0.305, but it is a single model with a documented "generalization wall": performance collapses on epitope clusters unlike the training set, and the model never looks at conservation or function. GEP shows that surface-mesh representations with Heat Kernel Signatures and equivariant graph networks capture complementary information for epitopes versus paratopes (MCC gains of 0.09 and 0.12 over non-geometric models on a sequence-dissimilar SabDab subset), but requires a paired antibody structure and models shape only. miTALOS is the only tool that already fuses ensemble target prediction with enrichment, but its targets are linear miRNA–mRNA pairs scored by five sequence tools, its enrichment runs downstream on gene lists with a network proximity score, and it has no structural or conservation awareness whatsoever. My claim is that moving the enrichment statistic inside the predictor — as a residue-level multiplicative gate on a multi-representation consensus — both breaks the generalization wall (because functional annotations are invariant to surface novelty) and specifically reranks predictions away from escapable variable loops and toward invariant functional cores.

Research Program

The method, which I will call FuncGate, has three components built around a residue $i$ on an antigen surface.

The first component is the predictor bank. For each residue I compute normalized scores from at least three orthogonal families: a surface-geometry score in the spirit of O-GEP (Heat Kernel Signatures on a mesh, capturing protrusion and solvent accessibility), a graph-geometry score in the spirit of I-GEP/WALLE (an equivariant network on the residue graph, capturing local contact topology), and a sequence score from protein language model embeddings plus a conservation score computed from a multiple alignment of the pathogen family (for coronaviruses, alignments in the style of the 44-genome Sarbecovirus comparison of SARS-CoV-2 gene content and COVID-19 mutation impact). The design principle, taken from CoBDock and Sen et al., is that the constituent scorers must fail differently: geometry errs on recessed sites, graphs err on novel topologies, and language models err on rare functions.

The second component is the functional gate. For each residue I take a structural neighborhood (all residues within a fixed angstrom radius of residue $i$'s side-chain centroid) and run a Fisher exact test of the neighborhood's residue set against a residue-level annotation universe: Gene Ontology Molecular Function terms mapped to specific residues (catalytic residues, binding residues), KEGG pathway membership of the antigen protein, and curated active-site annotations. The gate value is $F(i) = -\log_{10} p_i$. This is a direct, residue-level analogue of standard overrepresentation analysis, and it encodes the choke-point hypothesis quantitatively: a patch whose residues are statistically enriched for catalytic or binding annotations is a patch the pathogen cannot casually mutate — the same fitness-constraint logic that Hu et al. document for HIV sequence diversity and that Yang et al. and Amporndanai et al. exploit qualitatively when targeting coronavirus main proteases. It also addresses the incompleteness worry that motivates DeepSyn: where residue-level curation is missing, the gate degrades gracefully to no opinion rather than to a false negative, and I will test a literature-mined fallback using DeepSyn-style knowledge-graph associations at the protein level.

The third component is the combination. A small calibrator (logistic regression or a two-layer MLP, deliberately modest to avoid re-inventing WALLE) maps the predictor scores to a probability, and the gate enters multiplicatively:

$$E(i) \;=\; m\big(s_{\mathrm{surf}}(i),\, s_{\mathrm{graph}}(i),\, s_{\mathrm{seq}}(i),\, c(i)\big)\;\cdot\;\sigma\big(a + b\,F(i)\big),$$

where $c(i)$ is the alignment-derived conservation score, $\sigma$ is the logistic function, and $a, b$ are learned so that an unannotated patch ( $F(i)=0$ ) is not penalized below its consensus probability — enrichment can only lift a candidate, never silently delete it. This asymmetry is a deliberate response to the known failure mode of hard voting, documented across the ensemble-filtering literature from imiRTP-style multi-algorithm thresholds through the ceRNA pipelines of Zhang et al.: requiring agreement from every signal destroys the rare, non-canonical true positives that only one tool sees.

Beyond the method, the program includes a set of choke-point case studies for external, biological validation: the SARS-CoV-2 main protease catalytic pocket, the nucleocapsid RNA-binding groove shown nearly invariant across coronaviruses by Fan et al., the influenza HA stem/220-loop region that Chen et al. show blocks entry across H1N1, H3N2, and H9N2, and the conserved Region I of TNFα where Weng et al. designed sub-nanomolar cross-species minibinders.

Baselines

The naive reference point is conservation alone: rank residues by a Shannon-entropy conservation score from a pathogen-family alignment, take the top patches. This is what a virologist with no ML infrastructure would do, and it is the cheapest possible competitor; the proposal fails outright if FuncGate does not beat it on the case studies, because the whole point of adding structure and function is to fix conservation's known blindness to occlusion.

The strongest existing competitor is WALLE as benchmarked on AsEP, which reports an MCC of 0.305 on the 1,723-complex AsEP dataset — the best fast epitope predictor in the file. It will be run on exactly the same AsEP splits, with the same inputs and the same per-residue evaluation, and FuncGate must exceed its overall MCC and, more importantly, its MCC on the unseen-epitope-cluster split where the AsEP authors document the generalization wall. A secondary strong competitor is AlphaFold2-Multimer, which reports an MCC of 0.262 on AsEP at 1.66 hours per prediction; FuncGate must stay in the seconds-per-complex range to be usable at family scale, so the comparison is accuracy and wall-clock.

Two further comparisons isolate the proposal's key ingredients. First, a hard-voting ensemble of the same predictor bank (accept a residue if at least $k$ of the component scorers rank it highly, the imiRTP/PARCE convention) — this ablates the learned consensus against the field's default aggregation and directly tests whether multiplicative gating beats intersection. Second, the two-stage pipeline that a reviewer will say already exists: run WALLE, then post-hoc run standard GO/KEGG overrepresentation on the predicted patches without any feedback into prediction. This ablates the placement of enrichment (inside the score versus downstream) and is the single most important control in the proposal, because if post-hoc enrichment recovers the same choke points, the feedback loop adds nothing and the idea collapses to a workflow paper.

All baselines get the same data, the same splits, the same tuning budget (a fixed validation split of AsEP with identical hyperparameter search protocol), and the same compute cap per complex.

Why not just…?

Why not just use one strong end-to-end model? The file's evidence is blunt: on AsEP, EpiPred reports an MCC of 0.029 and MaSIF-site 0.037 — effectively no signal for antibody-specific epitopes — and even AlphaFold2-Multimer's respectable 0.262 costs 1.66 hours per prediction, which is unusable for scanning hundreds of variants across a pathogen family. Moreover, the AsEP ablations show that removing either the sequence-language or the structural-graph component of WALLE degrades performance substantially, which is itself an argument for ensembling representations rather than betting on one. The real threat here is a future single model trained on far more complexes; the cluster-split experiment addresses this by testing whether the functional gate's benefit persists even when the predictor bank is held fixed.

Why not just run standard overrepresentation analysis afterwards? This is the most serious alternative, and I treat it as an ablation rather than an argument. The literature gives two reasons to expect it falls short. Standard ORA suffers the "bag of genes" fallacy that miTALOS documents — unweighted set membership ignores where things sit relative to each other, which for epitopes means ignoring that a variable loop three residues from a catalytic site can sterically occlude it — and the discovery paradox that DeepSyn documents, where the most novel biology returns zero significant terms because it is not yet curated. But honestly: if a post-hoc enrichment filter applied to WALLE's output matches FuncGate's choke-point recovery on the case studies, the simpler pipeline wins and should be used. The two-stage ablation is designed to settle exactly this, and I will report it either way.

Why not just use sequence conservation? Because conservation of a residue says nothing about its accessibility or its structural consequences. Weng et al. provide the sharpest case in the file: the TNFα interface is broadly conserved between dog and human, yet a single non-conserved residue outside the core (Phe83 versus Ile83) displaces an adjacent loop and costs a clinical nanobody 1,000-fold in affinity — a fact invisible to any linear alignment score. Likewise, the HA 220-loop is strictly sequence-conserved yet immunologically silent on the native trimer because surrounding variable loops bury it, as the iterative immunogen study had to work around by grafting it onto heterologous scaffolds. Conservation alone also cannot tell a recessed choke point from a conserved but irrelevant buried helix; the geometry consensus and the surface predictor exist precisely to supply that missing axis.

Why not just majority-vote the predictors, as the field already does? Hard voting is the incumbent aggregation everywhere from imiRTP to PARCE, and it is cheap. Its documented cost is the sensitivity penalty: requiring $k$ votes deletes genuine interactions that only one specialized tool catches, and the threshold $k$ is picked empirically with no error control — both listed as open problems in the consensus literature. More decisively for this proposal, hard voting has no way to express "this site is conserved and catalytic even though only the graph model sees it." The multiplicative functional gate is exactly the mechanism that rescues such sites, and the hard-vote ablation will show whether that rescue matters in practice.

Proposed Experiments and Decisive Checks

The primary evaluation runs on AsEP (1,723 antibody-antigen complexes), with the cluster-based split the benchmark provides: train on some epitope clusters, test on unseen ones. The headline metric is per-residue MCC on the unseen-cluster split, compared against WALLE's reported 0.305 overall MCC and its (unreported-per-split) degraded value on unseen clusters; FuncGate's claim succeeds only if the unseen-cluster gap shrinks substantially, because the functional gate's annotations do not change when the surface does. Secondary metrics are AUC and top-$k$ patch recovery.

The second evaluation runs the GEP benchmark suite: the SabDab test subset of 49 complexes filtered to at most 70% pairwise sequence identity with training, and the Docking Benchmark v5 of 25 bound/unbound complexes, both from Geometric epitope and paratope prediction. This tests whether the surface-geometry and graph components of the bank reproduce known complementary behavior, using the paper's reported MCC improvements of 0.09 and 0.12 from Heat Kernel Signatures as the internal reference for what each representation contributes.

The third evaluation is the choke-point case studies, run as retrospective ranking tests. For each of the four systems (coronavirus main protease pocket, nucleocapsid RNA-binding groove, HA stem/220-loop, TNFα Region I), the question is whether FuncGate ranks the experimentally validated conserved site in its top few patches when given only the apo structure and a family alignment. Ground truth comes from the file: the protease covalent-inhibitor structures of Amporndanai et al., the cross-coronavirus nucleocapsid drug findings of Fan et al., the cross-subtype HA fusion blockade of Chen et al., and the TMB-1 minibinder affinities ($K_D$ of 0.55 nM on canine and 1.88 nM on human TNFα) of Weng et al.. The TNFα system doubles as a false-positive test: the method should not award choke-point status to the Phe83/Ile83 region, since a single substitution there is tolerated in the alignment yet lethal for binding — if conservation and function jointly misrank it, the gate's fitness-constraint logic is too coarse. A fifth, out-of-domain probe applies FuncGate to neutrophil elastase (PDB: 1ppg), the PARCE target, to check whether the method flags the evolved peptide's binding site without any antibody in the input.

The fourth check is cost. FuncGate must run in seconds per complex (the WALLE regime), because its use case is scanning a whole pathogen family's variants; any component approaching AlphaFold2-Multimer's 1.66 hours per prediction is disqualifying by design, and wall-clock will be reported alongside every accuracy number. No reported result found on Lacuna for runtime of the GEP models, so I will measure those myself on the same hardware as WALLE.

The decisive experiment for the whole proposal, however, is the two-stage ablation in the baselines section: FuncGate versus post-hoc standard ORA on identical predictor outputs, on the unseen-cluster AsEP split and on all four case studies. If the gate's placement inside the score buys nothing over enriching afterwards, the correct paper is a benchmarking study, not a method — and I would rather learn that in month six than in review.

References

  1. Cobdock: an accurate and practical machine learning-based consensus blind docking method
  2. Predicting binding sites of hydrolase-inhibitor complexes by combining several methods
  3. PARCE: Protocol for Amino acid Refinement through Computational Evolution
  4. AsEP: Benchmarking Deep Learning Methods for Antibody-specific Epitope Prediction
  5. Geometric epitope and paratope prediction
  6. miTALOS: Analyzing the tissue-specific regulation of signaling pathways by human and mouse microRNAs
  7. SARS-CoV-2 gene content and COVID-19 mutation impact by comparing 44 Sarbecovirus genomes
  8. Dissecting the Dynamics of HIV-1 Protein Sequence Diversity
  9. Structure-assisted design of inhibitors targeting coronavirus main proteases
  10. Inhibition mechanism of SARS-CoV-2 main protease by ebselen and its derivatives
  11. Deep functional synthesis: a machine learning approach to gene functional enrichment
  12. Genome-wide search for miRNA-target interactions in Arabidopsis thaliana with an integrated approach
  13. miRspongeR: an R/Bioconductor package for the identification and analysis of miRNA sponge interaction networks and modules
  14. The highly conserved RNA-binding specificity of nucleocapsid protein facilitates the identification of drugs with broad anti-coronavirus activity
  15. Asprellcosides B of Ilex asprella Inhibits Influenza A Virus Infection by Blocking the Hemagglutinin- Mediated Membrane Fusion
  16. Design of minibinder proteins as universal antagonists against canine and human TNFα
  17. Iterative immunogen optimization to focus immune responses on a conserved, subdominant viral epitope