Models that predict how strongly a small molecule binds a protein — the continuous quantity that decides whether a drug candidate is potent or useless — look impressively accurate on standard benchmarks, but much of that accuracy may be memorization in disguise. The cleanest demonstration comes from virtual screening: a 3D convolutional network whose protein atoms were replaced by a single dummy atom still achieved an AUC of 0.98 on the DUD-E benchmark, indistinguishable from the full model, and dropped to 0.50 — pure chance — once the decoy bias was removed (Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening). The model was reading the ligand, not the binding physics. In regression the same failure is quieter: standard train/test splits of PDBbind let models perform nearest-neighbor lookup over similar ligand scaffolds and homologous pockets rather than learning molecular recognition.
This proposal makes one specific bet: a model that genuinely computes binding physics must place its explanation — the attribution of which atoms and contacts drove the predicted energy — on the actual binding interface, and we can measure this automatically, without a human in the loop, and turn it into both a training signal and a data-selection signal. Explanation then becomes the connective tissue between three problems usually studied apart: what the model relies on (unified graph explanation), what data we show it (dataset diversity), and whether our evaluation can even detect shortcut reliance (benchmark contamination).
Core Claim
The claim has three parts. First, unified graph explainers — methods that jointly score node attributes and structural context rather than either in isolation, because an edge often matters only given the features it carries — have only ever been formulated and validated for discrete classification on synthetic motif benchmarks; extending layer-wise decomposition of the kind used in DEGREE: Decomposition Based Explanation for Graph Neural Networks to continuous affinity regression on 3D protein–ligand interaction graphs is open and requires new derivations (DEGREE needs an explicit decomposition operator per layer type, and geometric layers with distance-based Morse potentials have none). Second, this attribution gives an automated oracle for shortcut detection: non-covalent interaction profiles (as used in ArkDTA: attention regularization guided by non-covalent interactions for explainable drug-target binding affinity prediction) and RMSD-annotated decoy poses (as used in DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction) define, without any human annotation, which atoms the model should be using. Attribution mass landing elsewhere — on scaffold atoms far from the interface — is a machine-detectable shortcut. Third, this signal can be used actively, to select which training complexes matter, replacing the human expert that the closest prior work cannot do without.
The closest existing work is XIGL (Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance), which shows that querying graphs whose explanations the model is over-confident about, and penalizing attribution on human-marked spurious nodes, recovers out-of-distribution accuracy with under 20% of the annotations. It differs from this proposal in every load-bearing component: it uses first-order input gradients (which the unified-explanation literature documents as unfaithful and blind to feature–topology entanglement), it requires a human to draw ground-truth masks, and it is evaluated only on synthetic classification (its GIN baseline sat near 0% accuracy on unconfounded test data without guidance). We keep its active-learning insight and replace the gradient, the human, and the toy domain.
Research Program
The first component builds the explainer. We extend decomposition-based explanation to a geometric interaction graph neural network (GIGN, the regression head of the Folding-Docking-Affinity pipeline of Protein-ligand binding affinity prediction: Is 3D binding pose needed?), deriving per-layer decomposition operators for continuous regression outputs and distance-potential edge features. The output is an attribution $\phi_i \geq 0$ for each atom and interfacial contact, with $\sum_i \phi_i$ tied to the predicted energy, which also lets us reuse Shapley-style efficiency axioms from GraphSVX: Shapley Value Explanations for Graph Neural Networks as a sanity constraint. DEGREE's reported efficiency matters here: 0.991 explanation AUC on BA-Shapes at 0.44 seconds versus over 116 seconds for PGExplainer (DEGREE), because we will need to compute attributions for tens of thousands of complexes during active selection.
The second component converts the explainer into an automated teacher. Given a training complex, a PLIP-style non-covalent interaction profile defines the contact set $C$ of atoms that physically participate in binding. The training loss becomes
$$\mathcal{L} = \mathcal{L}_{\text{reg}} + \lambda \sum_{i \notin C} \phi_i^2,$$
where $\mathcal{L}_{\text{reg}}$ is the affinity regression loss, $\phi_i$ is the unified attribution of atom $i$, and the penalty pushes explanation mass off any atom that physics says is irrelevant. This is the automated replacement for XIGL's human-drawn masks. DecoyDB's 5.3 million decoy poses with continuous RMSD labels supply a second, richer oracle: as a ligand pose is displaced, the correct explanation should migrate with the actual contacts, so attribution-migration consistency across a decoy ladder becomes a self-supervised regularizer — a use of the decoy resource that DecoyDB itself, which operates as a black-box contrastive feature extractor, does not attempt.
The third component uses explanation statistics for active data selection. Following XIGL's Minimum Explanation Entropy strategy — query complexes where the model is most confidently and most spuriously explained — we select training complexes from the combined pool of PDBbind and DecoyDB crystals, targeting the model's own shortcuts rather than generic diversity. This is where dataset diversity enters as a mechanism, not a slogan: the face-space active learning result showed that representation-distance selection on under 20% of data beat baselines trained on everything (Actively Exploring Creation of Face Space(s) for Improved Face Recognition), and we test whether explanation-entropy selection, which conditions on what this model is exploiting, beats that model-agnostic criterion at matched budget.
The fourth component builds the evaluation that can actually falsify us. DyVal showed that dynamic, on-the-fly task generation exposes memorization that static benchmarks hide, but it is restricted to symbolic reasoning over directed acyclic graphs (DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks); the file's own gap analysis flags extending dynamic generation to 3D molecular structures as open. We build a molecular analog: dynamically generated test complexes via scaffold transplanting, pocket-preserving ligand mutation, and re-docking at controlled RMSD, so every evaluation instance is structurally novel and unmemorizable, with difficulty tuned continuously the way DyVal tunes DAG depth.
Proposed Experiments
All models share the GIGN backbone, identical compute, and an identical hyperparameter tuning budget per method; the only differences are the training signal and the data selection. Evaluation covers PDBbind v2016 and v2020 (RMSE and Pearson $R$ on the 2016 core set), CASF-2016 scoring power, the leakage-proof LP-PDBbind split from DecoyDB, the "both-new" zero-shot split on DAVIS from the FDA paper, and screening enrichment on both DUD-E and its bias-corrected Actives-as-Decoys variant from the DUD-E bias study. Attribution fidelity is scored as precision of top-$k$ attributed atoms against PLIP contacts. The activity-cliff derivative benchmark from PIGNet2 (where it reports an average Pearson $R$ of 0.64) is our sharpest test of the lead-optimization blindspot: ranking near-identical derivatives is exactly where scaffold memorization fails and contact physics must carry the prediction.
Baselines
The naive reference point is a ligand-only model, the regression analog of the DUD-E bias study's dummy-atom stress test; the file reports AUC 0.98 for it on DUD-E and 0.50 on the Actives-as-Decoys benchmark, but no reported result found on Lacuna for a ligand-only regression model on PDBbind or LP-PDBbind, so we will establish that number ourselves as the shortcut floor. The strongest existing competitor is a GIGN fine-tuned after DecoyDB contrastive pretraining, which reports an RMSE of 1.188 on the PDBbind 2016 core set (its best baseline, plain GIGN, reports 1.263) and is the current number to beat on leakage-proof evaluation (DecoyDB). We beat it only if we match its core-set RMSE within noise and improve clearly on LP-PDBbind, the dynamic decoy tests, and activity cliffs — winning on the contaminated split alone would count against us. The physics-informed baseline is PIGNet2, which reports $R = 0.747$ on CASF-2016 and $\text{EF}_{1\%} = 31.2$ on DUD-E against GNINA's 18.8 (PIGNet2); no reported result found on Lacuna for PIGNet2 on the Actives-as-Decoys benchmark, so its bias-robust screening power is unknown and we will measure it. Two ablations isolate the key ingredients: one swaps the unified decomposition attribution for first-order input gradients (XIGL-style) while keeping everything else, isolating the unified-explanation ingredient; the other keeps unified attribution but removes active selection in favor of random selection at the same budget, isolating the explanation-entropy selection ingredient.
Why not just…?
Why not just pretrain on DecoyDB and evaluate on LP-PDBbind? This is the real default, and we do not argue around it — we run it as our strongest baseline, and our method is deliberately stackable on top of the same pretraining. The honest expectation, grounded in the DUD-E bias study's finding that diagnostic ablations alone left deep models no better than classical docking once bias was removed, is that pretraining improves generalization without telling you what the model relies on; if our attribution-guided term and selection add nothing on LP-PDBbind beyond DecoyDB pretraining, the core claim is wrong and we will say so.
Why not just extend ArkDTA's non-covalent-interaction regularizer to 3D graphs? This is the most serious threat, since ArkDTA already couples affinity regression with interaction-profile supervision, and we include it head-to-head as an ablation: an auxiliary contact loss on the same GIGN backbone. It differs in two ways we expect to matter. Its supervision is binary contact presence, so it cannot express how much of the prediction each contact carries or detect a model that satisfies the auxiliary loss while still routing the regression through scaffold shortcuts; and it is a fixed training signal with no data-selection component. If the head-to-head shows the auxiliary loss matching our attribution penalty on all splits, we will report that the selection component must then carry the entire value of the proposal.
Why not just curate a diverse training subset by representation distance? The face-space result shows representation-distance selection is strong, and we run it at matched budget. But it is model-agnostic: it spreads the data without knowing which shortcut this model has adopted. The XIGL result suggests shortcut usage is model-specific, and the DUD-E study shows molecular shortcuts are subtle joint motif–scaffold effects that generic coverage metrics can miss. The matched-budget comparison is exactly the experiment that settles this; if distance-based curation ties, the explanation-entropy mechanism is not earning its complexity.
Why not just rely on leakage-proof splits and skip dynamic evaluation? LP-PDBbind filters structural and ligand overlap at construction time, but continuous contamination — pocket-shape similarity and scaffold analogy below any discrete threshold — survives static filtering, just as models scoring high on static reasoning benchmarks collapsed to near zero on dynamically generated tasks in DyVal, with GPT-4's reachability accuracy alone dropping 13.6% under a mere description-order reversal. Our dynamic decoy tests are the only component that can catch shortcuts neither we nor the split designer anticipated; they are the falsification instrument, not decoration.
Decisive Checks
The proposal stands or falls on three numbers. On LP-PDBbind and the dynamic decoy suites, the explanation-guided model must beat the DecoyDB-pretrained baseline (whose core-set RMSE of 1.188 defines the standard) while a ligand-only ablation of the same trained model shows near-zero retained performance — meaning the model genuinely reads the pocket. On activity cliffs, it must beat PIGNet2's $R = 0.64$, since derivative ranking is impossible by scaffold memorization. And on the matched-budget selection curves, explanation-entropy selection must beat both random and representation-distance selection, echoing the under-20%-of-data efficiency that the face-space and XIGL results each independently demonstrated in their domains. If the first two hold and the third fails, the useful residue is a validated unified explainer for continuous molecular regression — itself a gap the file identifies as completely open — and we would write the proposal's sequel around that.
References
- Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening
- DEGREE: Decomposition Based Explanation for Graph Neural Networks
- ArkDTA: attention regularization guided by non-covalent interactions for explainable drug-target binding affinity prediction
- DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction
- Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance
- Protein-ligand binding affinity prediction: Is 3D binding pose needed?
- GraphSVX: Shapley Value Explanations for Graph Neural Networks
- Actively Exploring Creation of Face Space(s) for Improved Face Recognition
- DyVal: Dynamic Evaluation of Large Language Models for Reasoning Tasks
- PIGNet2: a versatile deep learning-based protein–ligand interaction prediction model for binding affinity scoring and virtual screening