Debiasing Hub Residues: Popularity-Bias Correction for Allosteric Mutation Prediction

Proteins are not democratic: a handful of tightly packed core residues touch dozens of neighbours, while the flexible surface loops where allosteric regulation so often happens touch only a few. When we feed a protein's residue contact graph to a graph neural network — a model that builds each residue's representation by repeatedly averaging in its neighbours' — the well-connected core residues end up shaping everyone else's embeddings far more than the sparsely connected periphery. Recommender systems researchers have studied exactly this failure mode for years under the name popularity bias: message passing amplifies high-degree "hub" nodes, their embedding vectors grow disproportionately large, and low-degree "tail" items become nearly invisible to the model. The mathematics is worked out — one proof shows an item's influence on training gradients scales as $d_i^{3/2}$, where $d_i$ is its degree, even under the standard degree normalization that is supposed to prevent it (How Graph Convolutions Amplify Popularity Bias for Recommendation?).

This proposal argues that the same pathology afflicts structural protein models, and that it lands exactly where allosteric prediction can least afford it. Allostery is the mechanism by which a perturbation at one site of a protein changes the behaviour of a distant functional site, often more than 10–30 Å away, frequently through low-contact, flexible residues rather than the rigid core (From Binding-Induced Dynamic Effects in SH3 Structures to Evolutionary Conserved Sectors). If graph models over-weight high-degree core residues and under-represent the sparse periphery, they will systematically misjudge distal mutations — the "local blindspot" already documented for sequence- and structure-based tools in cancer gene panels (Understanding the impacts of missense mutations on structures and functions of human cancer-related genes: A preliminary computational analysis of the COSMIC Cancer Gene Census). The plan is to first measure whether this hub bias exists in protein graphs, then correct it by transferring three mature popularity-bias mitigation techniques, and finally test whether the correction specifically improves prediction of distal, allosteric effects without degrading overall accuracy.

Core Claim

Graph neural networks that message-pass over protein residue contact graphs inherit the degree-amplification pathology proven for recommender graphs, and this pathology is a concrete, correctable cause of the known weakness of variant-effect models on distal allosteric mutations. Correcting it requires no new architecture: degree-aware normalization of message passing (in the spirit of RawlsGCN: Towards Rawlsian Difference Principle on Graph Convolutional Network), post-hoc projection of a learned "hub direction" out of residue embeddings (adapting Post-hoc Popularity Bias Correction in GNN-based Collaborative Filtering), and adaptive temperature control in contrastive alignment (from Adap-τ : Adaptively Modulating Embedding Magnitude for Recommendation) should each, and jointly, improve prediction on low-degree and distally located mutations while leaving head-region performance intact.

The proposal is honest that the transfer is a hypothesis, not a derivation. The $d_i^{3/2}$ result was proven for user–item bipartite graphs; protein contact graphs have different topology and training signal, so Phase 1 is a diagnostic measurement before any correction is applied.

Research Program

The program has three phases, each producing a publishable result regardless of whether the hypothesis survives.

Phase 1 — Diagnose. We build a supervised variant-effect predictor: a message-passing GNN over the AlphaFold2-derived residue contact graph of each protein, initialized with ESM-2 sequence embeddings, trained on the ProteinGym substitution benchmark — 217 deep mutational scanning assays covering roughly 696,000 single amino acid substitutions, with labels binarized into the top and bottom 30% per assay, exactly the protocol of TriFit: Trimodal Fusion with Protein Dynamics for Mutation Fitness Prediction. We then stratify test mutations two ways: by the contact degree of the mutated residue (quartiles), and by its distance to the annotated active or binding site (near vs. distal, threshold at 10 Å, matching the allostery literature's definition of remote). The diagnostic questions are whether AUROC degrades on the low-degree and distal strata, and whether embedding norm correlates with residue degree — the protein-graph analogue of the finding that item embedding magnitude grows with training frequency in Adap-τ. If neither bias appears, the hub-bias hypothesis is falsified for this regime and the program pivots to the database-level skew documented in ASD2023: towards the integrating landscapes of allosteric knowledgebase (where a few well-studied families dominate allosteric annotations), which is itself a popularity-bias problem, just at the dataset rather than the architecture level.

Phase 2 — Correct. Three interventions, each drawn from a distinct mitigation family and each ablatable. First, in-training degree normalization: replacing symmetric normalization with a doubly stochastic or Rawls-style scheme (RawlsGCN) so no residue's vote in message passing scales with its contact count. Second, post-hoc projection: after training an unmodified model, estimate a "hub direction" $\hat{v}$ as the normalized difference between the centroid embedding of high-degree residues and that of low-degree residues, then subtract its component from every residue embedding,

$$h'_i = h_i - (h_i \cdot \hat{v})\,\hat{v},$$

leaving the component of the representation orthogonal to hub-ness — a direct transplant of the technique that yielded the largest tail gains in recommender graphs (Post-hoc Popularity Bias Correction). Third, adaptive temperature: where the model uses contrastive alignment (as TriFit does with InfoNCE), we replace fixed temperature with the closed-form global $\tau_0$ plus per-context adaptation of Adap-τ. This is motivated by the proof in Contrastive Learning for Debiased Candidate Generation in Large-Scale Recommender Systems that InfoNCE with data-distributed negatives acts as implicit inverse propensity weighting — meaning TriFit's unweighted negative sampling may already be debiasing, but in an uncontrolled way whose strength depends on the temperature it never tunes.

Phase 3 — Validate on allosteric-specific tasks. Overall AUROC on ProteinGym can hide a distal-stratum failure, so the decisive tests are allosteric ones. We evaluate whether debiased residue embeddings improve allosteric site identification on the Allosite benchmark — 90 non-redundant crystal structures from the ASD, where a support-vector-machine classifier on pocket descriptors reached greater than 92% cross-validation accuracy and greater than 95% on an independent test set (Allosite: a method for predicting allosteric sites) — and we test whether post-hoc projection applied to a pre-trained dynamic pathway model recovers force-transmission pathways without AlloPool's expensive iterative edge pruning, using its Pin-1, ADGRG1 GAIN, and B1AR trajectory benchmarks (AlloPool: An Adaptive Graph Neural Network for Dynamic Allosteric Network Prediction in Protein Systems). The biological grounding comes from work showing that distal "stiff cores" and hub residues are genuine physical conduits in SOD1 (Molecular mechanisms underlying the impact of mutations in SOD1 on its conformational properties associated with amyotrophic lateral sclerosis as revealed with molecular modelling) and ARID domains (Communication Routes in ARID Domains between Distal Residues in Helix 5 and the DNA-Binding Loops) — the point is not that hubs are unimportant, but that a model should represent them accurately rather than amplify them as a graph-convolution artifact.

Baselines

The "do nothing" reference point is the raw message-passing GNN of Phase 1 with standard symmetric degree normalization and no debiasing: same architecture, same ESM-2 initialization, same ProteinGym training. This baseline doubles as the ablation that isolates the proposal's key ingredient, because every intervention is applied on top of it. The no-graph reference is ESM3 used zero-shot, which reports an AUROC of 0.769 on the ProteinGym substitution benchmark (as reported in TriFit); it tells us how much of the distal-stratum gap is already closed by a large protein language model that never sees a contact graph.

The strongest existing competitor is TriFit (TriFit), which fuses sequence, static structure, and coarse-grained dynamics features and reports AUROC 0.897 with an Expected Calibration Error of 0.044 on the same benchmark and binarization we use. Other reported reference points from the same evaluation are Kermut at 0.864 and ProteinNPT at 0.844. TriFit is the right strongest baseline because it currently owns the benchmark we target and because its dynamics modality — GNM flexibility and cross-correlations — is precisely the signal type that hub-dominated message passing could distort; beating it on the distal stratum while matching it overall is the proposal's central claim made concrete. To keep the comparison fair we rerun TriFit's released model and code on our identical assay splits and binarization, give every method the same hyperparameter tuning budget (one matched grid, same number of seeds; TriFit reports training stability with a standard deviation of 0.0002, so differences below that are noise), and give all methods the same structures from the same structure source.

Additionally, a plain inverse-propensity-weighted variant of our GNN — training loss reweighted by inverse residue degree or inverse assay frequency — serves as the cheap debiasing baseline a reviewer would demand; it is described in the next section. Beating the baselines means: overall AUROC statistically above TriFit's 0.897, or within its noise band, and at least 2 AUROC points better than every baseline on the pre-registered low-degree and distal strata, with calibration (ECE) not worse than TriFit's 0.044. On the Allosite independent test we compare against the reported greater-than-95% accuracy under their split protocol; on the Pin-1 Apo trajectory we compare pathway-recovery fidelity against AlloPool's reported RMSD of 1.49 Å and the Neural Relational Inference baseline's 4.83 Å.

Why not just…?

Why not just use a bigger protein language model? Because the evidence in the file says sequence models miss exactly the signal we target: within TriFit's own ablations, adding biophysical dynamics to a sequence-plus-structure model gained 2.2 AUROC points, while adding static structure to a sequence-only model gained only 0.2, and zero-shot ESM3 sits at 0.769 — all reported in TriFit. Coarse-grained dynamics features carry the allosteric coupling that pure sequence coevolution does not. That said, this is a real threat: a pLM trained on hundreds of millions of sequences may absorb coevolutionary signals our contact-graph model lacks. We settle it empirically — ESM3 zero-shot is a baseline on every stratum, and if it closes the distal gap by itself, our diagnostic will show it and the debiasing interventions then apply on top of a sequence model rather than a graph model, which is a finding either way.

Why not just use a standard GNN with the usual degree normalization? Because the normalization does not work, in the recommender setting at least: How Graph Convolutions Amplify Popularity Bias for Recommendation? proved that gradient influence still scales as $d_i^{3/2}$ despite symmetric normalization, and Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System found graph-based models more vulnerable to popularity bias than plain matrix factorization. The honest caveat is that both results are for bipartite interaction graphs, and protein contact graphs might not inherit the pathology — which is why the unmodified normalized GNN is our "do nothing" baseline and why Phase 1 measures the bias before we claim to fix it.

Why not just reweight the training loss by inverse frequency? Explicit inverse propensity weighting is the obvious cheap fix, and it might work; we run it as a baseline rather than argue it away. But the file documents repeated failures of this family: MACR's authors note IPW is difficult to tune and often unstable (Model-Agnostic Counterfactual Reasoning), the contrastive-generation work showed InfoNCE achieves implicit propensity weighting without brittle explicit propensity estimates (Contrastive Learning for Debiased Candidate Generation), and the post-hoc work found reweighting distorts representations early in training (Post-hoc Popularity Bias Correction). A related shortcut — normalizing embeddings to a fixed sphere with a hand-picked temperature — is also covered by evidence: Adap-τ showed fixed temperatures are hypersensitive and that its adaptive scheme beat exhaustive manual grid search by an average of 9% across benchmarks.

Why not just filter out hub residues or use a tight distance cutoff on edges? Because the hubs are the signal, not noise. Elastic-network modelling identified the stiff core of SOD1 as the very conduit that transmits destabilizing distal mutations (Molecular mechanisms underlying the impact of mutations in SOD1), and AlloPool showed static distance cutoffs and fully connected graphs fail to separate functional force conduits from inert contacts — its adaptive pruning was needed to reach 1.49 Å reconstruction on Pin-1 where a static-graph baseline reached only 4.83 Å. Deleting hubs would amputate the transmission pathway; the goal is to stop the model from amplifying them and instead represent them accurately.

Proposed Experiments and Decisive Checks

Experiment 1 is the diagnostic on ProteinGym (ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design): train the GNN, report AUROC by degree quartile and by near/distal split, and the Spearman correlation between embedding norm and degree. The decisive outcome is a reproducible gap between the top and bottom degree quartiles or between near and distal mutations; its absence redirects the program to database-level family skew in ASD2023, where the over-representation of kinases and GPCRs among annotated allosteric sites is the popularity-bias problem in its plainest form. Experiment 2 applies each intervention separately and jointly, with the pre-registered success criterion stated in the baselines section: match or beat TriFit's 0.897 overall AUROC while gaining at least 2 AUROC points on the low-degree and distal strata, and keep ECE at or below 0.044. Experiment 3 tests transfer: a pocket classifier on debiased residue embeddings evaluated on the Allosite 90-structure protocol against the reported greater-than-95% independent-test accuracy (Allosite), with a specific sub-analysis of allosteric pockets far from the orthosteric site, where the bias should hurt most. Experiment 4 applies the post-hoc hub-direction projection to a pre-trained dynamic pathway model on the Pin-1, ADGRG1 GAIN, and B1AR trajectories, comparing recovered pathway residues and reconstruction fidelity against AlloPool's reported 1.49 Å on Pin-1 Apo and the 4.83 Å of the Neural Relational Inference baseline; success here would mean distal-pathway recovery without re-running the expensive iterative pipeline, the same no-retraining advantage that made post-hoc projection attractive in recommender graphs. Throughout, the interpretive guardrail is the same: any gain must come from the tail strata without collapsing head performance, because the known accuracy-diversity trade-off in popularity-bias mitigation warns that aggressive suppression of hub signal can destroy the cases where hub-ness genuinely reflects physical importance — in proteins, unlike in recommender systems, the hubs sometimes really are carrying the signal.