Paper deep dive

MPGE: A Multi-Perspective Graph Explainer for Molecular Classification Explanation

Mahtab Sarvmaili

Year: 2026Venue: arXiv preprintArea: cs.LGType: PreprintEmbeddings: 93

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 10/10/2026, 2:56:48 AM

Summary

The paper introduces MPGE (Multi-Perspective Graph Explainer), a framework for explaining molecular classification decisions made by Graph Neural Networks (GNNs). MPGE unifies three explanation perspectives: Factual (PT), which identifies compact edge subsets preserving the prediction; Counterfactual (CF), which finds minimal deletions that reverse the prediction; and Exemplar (EXE), which finds bounded deletions that preserve the prediction while maximizing structural difference. The method extends CF-GNNExplainer to graph classification, using a frozen GCN predictor and optimizing separate symmetric edge masks for each view. It includes a chemistry-aware fragment backend using RDKit to ensure chemical validity. Evaluation on datasets like MUTAG, Mutagenicity, AIDS, COX2_MD, and BBBP demonstrates coverage, stability, and runtime metrics.

Entities (13)

MPGE · method · 98%AIDS · dataset · 95%BBBP · dataset · 95%COX2_MD · dataset · 95%GNN · model · 95%MUTAG · dataset · 95%Mutagenicity · dataset · 95%RDKit · software · 93%GCN · model · 92%CF-GNNExplainer · method · 90%Counterfactual Explanation · concept · 90%Exemplar Explanation · concept · 90%Factual Explanation · concept · 90%

Relation Signals (11)

MPGE → evaluateson → COX2_MD

confidence 95% · We evaluate the primary GCN implementation on... COX2_MD...

MPGE → evaluateson → BBBP

confidence 95% · We evaluate the primary GCN implementation on... BBBP

MPGE → evaluateson → MUTAG

confidence 95% · We evaluate the primary GCN implementation on MUTAG...

MPGE → evaluateson → Mutagenicity

confidence 95% · We evaluate the primary GCN implementation on... Mutagenicity...

MPGE → evaluateson → AIDS

confidence 95% · We evaluate the primary GCN implementation on... AIDS...

MPGE → uses → GNN

confidence 95% · MPGE unifies factual support, counterfactual sensitivity, and exemplar tolerance for a frozen classifier [GNN].

MPGE → usesbackend → RDKit

confidence 93% · A separate BBBP fragment backend returns RDKit-sanitized molecules.

MPGE → extends → CF-GNNExplainer

confidence 92% · Our graph-classification extension of CF-GNNExplainer learns symmetric edge rankings

MPGE → includescomponent →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Graph neural networks (GNNs) predict molecular properties from chemical graph data, but predictive accuracy does not explain how graph information supports an individual decision. A compact prediction-preserving rationale does not necessarily reveal which changes reverse the decision or which modifications the model tolerates. We propose the Multi-Perspective Graph Explainer (MPGE), unifying factual support, counterfactual sensitivity, and exemplar tolerance for a frozen classifier. The factual view, originally termed prototype (PT), seeks a compact retained edge set with the same label and required confidence. Counterfactual (CF) explanations seek bounded prediction-changing deletions; exemplar (EXE) explanations seek non-trivial bounded deletions that preserve the label and confidence. A shared constrained formulation connects prediction behavior, compactness, and edit cost, while separate objectives generate the three views. Our graph-classification extension of CF-GNNExplainer learns symmetric edge rankings and verifies discrete candidates, recording unsuccessful searches. A separate BBBP fragment backend returns RDKit-sanitized molecules. We evaluate the primary GCN implementation on MUTAG, Mutagenicity, AIDS, COX2_MD, and BBBP using semantic coverage, conditional quality, stability, and runtime. Successful factual masks retained 8.6%--15.5% of input edges on average across datasets; bounded counterfactual coverage was 4.8%--67.6%, and exemplar preservation coverage was 98.9%--100.0%. Exploratory controls reveal the influence of hard projection and retained node information. Quantitative comparisons and molecular visualizations characterize model support, sensitivity, and tolerance without treating them as validated chemical mechanisms.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

89,551 characters extracted from source content.

Expand or collapse full text

MPGE: A Multi-Perspective Graph Explainer for Molecular Classification Explanation Mahtab Sarvmaili Affiliation: Dalhousie University Halifax, Nova Scotia, Canada Abstract Graph neural networks (GNNs) predict molecular properties from chemical graph data, but predictive accuracy does not explain how graph information supports an individual decision. A compact prediction-preserving rationale does not necessarily reveal which changes reverse the decision or which modifications the model tolerates. We propose the Multi-Perspective Graph Explainer (MPGE), unifying factual support, counterfactual sensitivity, and exemplar tolerance for a frozen classifier. The factual view, originally termed prototype (PT), seeks a compact retained edge set with the same label and required confidence. Counterfactual (CF) explanations seek bounded prediction-changing deletions; exemplar (EXE) explanations seek non-trivial bounded deletions that preserve the label and confidence. A shared constrained formulation connects prediction behavior, compactness, and edit cost, while separate objectives generate the three views. Our graph-classification extension of CF-GNNExplainer learns symmetric edge rankings and verifies discrete candidates, recording unsuccessful searches. A separate BBBP fragment backend returns RDKit-sanitized molecules. We evaluate the primary GCN implementation on MUTAG, Mutagenicity, AIDS, COX2_MD, and BBBP using semantic coverage, conditional quality, stability, and runtime. Successful factual masks retained 8.6%–15.5% of input edges on average across datasets; bounded counterfactual coverage was 4.8–67.6% and exemplar preservation coverage was 98.9–100.0%. Exploratory controls reveal the influence of hard projection and retained node information. Quantitative comparisons and molecular visualizations characterize model support, sensitivity, and tolerance without treating them as validated chemical mechanisms. Keywords: graph neural networks, explainable artificial intelligence, counterfactual explanations, exemplar explanations, molecular property prediction I Introduction Molecular property prediction connects chemical structure to outcomes such as mutagenicity, biological activity, and blood–brain barrier penetration. Graph neural networks (GNNs) provide a natural representation-learning approach to these tasks by processing atom features and their structural relationships [1]. Predictive performance alone, however, does not reveal which information supports an individual model decision. Explanations can help inspect that decision and formulate structural hypotheses, although model-faithfulness evidence must be distinguished from validated chemical knowledge [2]. Understanding a molecular prediction requires more than identifying one compact rationale. Factual explainers seek a subset of graph information that supports the current prediction [3, 4]. Such a subset can be sufficient without identifying the changes needed to reverse the decision. Counterfactual explainers instead search for prediction-changing perturbations [5, 6]. A successful counterfactual identifies sensitivity to a particular edit, but does not characterize which other changes the model tolerates. These distinctions motivate three complementary questions about the same query: what supports the prediction, what changes it, and what can change while the prediction is preserved? We address these questions through factual, counterfactual (CF), and exemplar (EXE) explanations. The factual view seeks a compact retained edge set that preserves the query label and required confidence. This view was originally termed prototype (PT) in our study; we use factual in the formal definition and experiments because it explains an individual query rather than a class-level prototype. The counterfactual view seeks a bounded deletion that changes the predicted property class. The exemplar view seeks a non-trivially modified query that retains its label and a prescribed fraction of its confidence. Here, an exemplar is a generated query variant, not a representative training example. A positive edit requirement excludes the unchanged query, while the budget restricts the permitted deletions. Its same-outcome-under-change condition is related to the neighborhood and prediction relations of Explanation Sets [7]. Together, the views describe prediction support, sensitivity, and tolerance within the chosen intervention space; they do not certify a chemical mechanism or robustness to arbitrary modifications. We propose the Multi-Perspective Graph Explainer (MPGE), which unifies these three views through a constrained formulation and a common explanation interface. Their objectives connect prediction preservation or change with compactness and edit cost. Prior work already couples factual and counterfactual reasoning [8, 6]; our focus is to include the prediction-preserving edited query and evaluate all three views under explicit output constraints. The structural implementation, MPGX-CF, extends the CF-GNNExplainer architecture to graph classification [5]. It shares a frozen predictor across perspectives but optimizes a separate symmetric edge mask for each. A ranked hard-mask search checks the actual returned graph and records a failure when no searched candidate meets the requested constraints. A separate BBBP fragment-search backend returns RDKit-sanitized molecules, distinguishing chemical representability from the structural feasibility of an edge mask. The contributions are: • a three-view formulation connecting compact factual support, bounded counterfactual change, and non-trivial exemplar preservation; • a graph-classification extension of CF-GNNExplainer with view-specific objectives, discrete feasibility checks, and explicit search-failure reporting; • a chemistry-aware fragment-search instantiation that evaluates the three views in a separately sanitized molecular candidate space; and • a multi-dataset evaluation that separates semantic coverage, conditional quality, and mask stability, with exploratory controls for discretization, retained node information, and edit budgets. We evaluate the primary GCN implementation on four TU molecular graph benchmarks—MUTAG, Mutagenicity, AIDS, and COX2_MD—and the scaffold-held-out BBBP dataset. Quantitative comparisons assess the three views using compatible baselines, while visualizations show how their retained and removed structures answer the different explanation questions. The following sections position the framework in the literature, define its objectives and search procedures, and present the evaluation together with the limits of molecular interpretation. I Related Work I-A Factual graph explanations GNNExplainer learns compact structural and feature masks that preserve information about a model prediction [3]. PGExplainer amortizes mask generation over multiple inputs and also includes size and entropy regularization [4]. These approaches motivate our differentiable edge ranking, but our evaluation treats the final hard subgraph, rather than its continuous surrogate, as the explanatory object. I-B Counterfactual and coupled reasoning CF-GNNExplainer freezes a trained GNN and learns a symmetric adjacency sparsification mask, retaining the best discrete deletion candidate that changes a node prediction [5]. We adapt this separation between predictor and perturbation parameters to graph classification. RCExplainer learns common decision logic and seeks complementary conditions on a selected subgraph and its remainder [6]. CF2 directly couples factual sufficiency and counterfactual necessity through a sparse relaxed mask [8]. These approaches establish coupled reasoning as an important precedent; our formulation additionally specifies a bounded, non-trivially edited query that preserves the prediction. Recent work formalizes post-generation graph-counterfactual minimization, proves its computational hardness, and proposes a bounded local search [9]. Our ranked-prefix scan is more restricted; it audits feasibility along one learned edge order and does not supply a global minimality guarantee. CLEAR learns a generative graph-level counterfactual model, while D4Explainer uses discrete denoising diffusion for distribution-aware counterfactual and model-level explanations [10, 11]. Their distribution objectives address a different question from whether a discrete graph changes a prediction. MPGE’s hard checks establish the latter; its separate fragment backend adds sanitation, not a learned molecular-manifold constraint. I-C Molecular and distribution-aware explanations Molecular counterfactual methods make this domain constraint concrete. MEG trains a reinforcement-learning generator over valid molecular actions, whereas MMACE searches model-agnostically through SELFIES mutations [12, 13]. More recent counterfactual masking replaces removed substructures with generatively completed fragments to reduce out-of-distribution evaluation artifacts [14]. ICExplainer likewise uses a learned graph distribution, and ATEX-CF (ICLR 2026) expands node-level counterfactual search to both edge additions and deletions [15, 16]. Our deletion-only fragment search explores a restricted candidate family. It instantiates all three views under one frozen predictor without claiming the broader generation capabilities of these methods. Recent work also changes the explanatory object or its optimization space. MAGE generates model-level graph explanations using motifs, including molecular examples [17]. IDEA aligns instance-level explanations in a prototypical representation space rather than only matching graph labels [18]. Neither a class-level generated motif nor a latent prototype is equivalent to the compact query-specific factual mask termed PT in the original version of our study. This distinction matters because prediction preservation alone does not identify a unique explanatory structure. I-D Exemplars and evaluation The same-outcome-under-change perspective motivates the invariance constraint used by EXE. Explanation Sets formalize several example-based views within a common neighborhood and prediction relation [7]. We adapt that structure to a generated subgraph and require a positive edit cost, preventing the original graph from being returned as its own exemplar. This definition is stated explicitly because other explainability work may use exemplar for a representative observed case rather than a transformed query. GraphFramEx separates explanation user needs and combines necessity and sufficiency evidence [19]. Other systematic analyses show that faithfulness does not imply stability [20], and explanation masks can be fragile to small perturbations [21]. We consequently report semantic success, coverage, size or edit cost, stability, and runtime separately instead of collapsing them into one unqualified accuracy score. Recent work additionally constructs degenerate GNN explanations that standard faithfulness metrics can miss [22]. Although that analysis targets self-explainable GNNs rather than our post-hoc setting, it motivates interpreting favorable fidelity together with structure and stability diagnostics rather than as a complete explanation-quality certificate. Accordingly, MPGE is positioned as a query-level, three-view formulation with explicit hard-output evaluation, not as the first combination of explanation perspectives or a distribution-certified molecular generator. I Problem Formulation Let G=(V,E,X)G=(V,E,X) be an attributed undirected graph and let a trained, frozen binary classifier f return a logit. Define p(G)=σ(f(G)),y(G)=[p(G)≥0.5].p(G)=σ(f(G)), y(G)=I[p(G)≥ 0.5]. (1) For the query graph G, fix y=y⁡(G)y=y(G) and define, for any candidate H, py(H)=p(H)p_y(H)=p(H) if y=1y=1 and py(H)=1−p⁡(H)p_y(H)=1-p(H) otherwise. Thus pyp_y always refers to the original query class, even after a flip. An edit set Δ induces T⁡(G,Δ)T(G, ). Structural candidates have H=(V,EH,X)H=(V,E_H,X) with EH⊆E_H E; molecular candidates are generated connected fragments. Structural validity requires symmetric deletion of existing edges with unchanged nodes and features. Molecular validity additionally requires the fragment-generation and RDKit sanitation checks. Neither predicate establishes distributional plausibility or a physical property. The evaluated structural queries have M=|E|≥1M=|E|≥ 1. Let AF(H)A_F(H) be a backend-specific factual size, C⁡(Δ)C( ) the edit cost, and BD_B the permitted bounded edits. Structurally, AF(H)=|EH|A_F(H)=|E_H|, C⁡(Δ)=|E∖EH|C( )=|E E_H|, and B=Δ:1≤C⁡(Δ)≤BD_B=\ :1≤ C( )≤ B\. The fragment backend uses retained heavy atoms for AFA_F, deletion steps for C, and an additional removed-heavy-atom fraction limit in BD_B. I-A Factual explanation A factual explanation is a compact retained graph SF⊆GS_F G. For the structural backend, SF=(V,EF,X)S_F=(V,E_F,X) retains every query node and feature and selects only EF⊆E_F E; its compactness is therefore edge compactness. The chemistry-aware backend instead returns a sanitized connected fragment under the same prediction constraint but a different size cost: minSF _S_F AF(SF) A_F(S_F) (2) s.t. .t. y⁡(SF)=y, y(S_F)=y, py(SF)≥τF, p_y(S_F)≥ _F, SF∈F(G),Valid(SF)=1. S_F _F(G), (S_F)=1. Here F(G)S_F(G) contains non-empty retained-edge graphs for the structural backend and generated connected fragments with at least two atoms for the molecular backend. The former excludes a zero-edge output but retains all node information. It also contains the full query, so factual feasibility is expected for a valid non-empty structural query; compactness, rather than coverage alone, distinguishes factual outputs. The confidence thresholds are relative to the unmasked prediction, τF=τE=ρpy(G) _F= _E=ρ p_y(G), with ρ fixed before test evaluation. I-B Counterfactual explanation A counterfactual is a minimum-cost valid intervention that changes the decision: minΔC _ _C C⁡(ΔC) C( _C) (3) s.t. .t. y(T(G,ΔC))≠y,ΔC∈B, y(T(G, _C))≠ y, _C _B, Valid⁡(T⁡(G,ΔC))=1. (T(G, _C))=1. If m is a keep mask, the deletion cost is proportional to ∥1−m∥1 1-m _1, not ∥m∥1 m _1. This direction is essential: penalizing the keep mask would reward deleting more edges. I-C Exemplar explanation An exemplar is a non-trivial, bounded change under which the decision and a minimum confidence are preserved: maxΔE _ _E U⁡(G,T⁡(G,ΔE)) U(G,T(G, _E)) (4) s.t. .t. y⁡(T⁡(G,ΔE))=y, y(T(G, _E))=y, py(T⁡(G,ΔE))≥τE, p_y(T(G, _E))≥ _E, ΔE∈B, _E _B, Valid⁡(T⁡(G,ΔE))=1. (T(G, _E))=1. Here U measures a specified structural difference from the query. Positive cost excludes a no-op and the budget restricts the edit family; neither guarantees meaningfulness to a domain expert. We use removed-edge fraction for masks and one minus Morgan-fingerprint similarity for molecular fragments. Factual and exemplar views share a prediction-preservation condition but optimize different quantities under different candidate constraints. The implementation searches restricted families and does not certify global optima. IV MPGE Framework and MPGX-CF Architecture IV-A Frozen predictor and perturbation generator Let fWf_W be the trained graph classifier and let gPg_P be its perturbation counterpart. As in CF-GNNExplainer [5], gPg_P uses the frozen weights W of fWf_W and learns only structural gates P. The predictor is in evaluation mode: dropout is disabled and BatchNorm running statistics are fixed. Our implementation assigns one logit zez_e to each of the M unique undirected edges and mirrors me=σ⁡(ze)m_e=σ(z_e) onto both directed message-passing entries. Consequently, A~=A⊙m A=A m is symmetric, cannot introduce a non-edge, and does not gate the self-messages added by the GCN. The GCN normalization is recomputed on A~+I A+I during every forward pass. This is the graph-classification analogue of the CF-GNNExplainer generating model, rather than a second predictor trained from data. The gate logits are initialized as ze=1+ϵez_e=1+ _e, with ϵe∼⁡(0,0.012) _e (0,0.01^2), giving initial soft weights near σ⁡(1)σ(1) rather than unit weights. Let qy(m)q_y(m) denote the probability assigned by gPg_P to the query class under the soft mask, m¯=M−1∑eme m=M^-1 _em_e, and H⁡(m)H(m) denote mean binary entropy. We optimize a separate P for each requested view; sharing the predictor does not imply sharing an explanation mask. IV-B Perspective-specific objectives The factual relaxation is LF=λp[τF−qy(m)]++β1m¯+β2m2¯,L_F= _p[ _F-q_y(m)]_++ _1 m+ _2 m^2, (5) where m¯ m is the mean retained mass. The counterfactual relaxation reverses the decision condition and penalizes removed rather than retained mass: LC=λp[qy(m)−0.49]++β1(1−m)¯+β2(1−m)¯2.L_C= _p[q_y(m)-0.49]_++ _1 (1-m)+ _2 (1-m)^2. (6) The exemplar relaxation preserves confidence while encouraging a non-trivial deletion up to budget b: LE=λp[τE−qy(m)]+−β1(1−m)¯+λb[(1−m)¯−b]+2+β2H(m). splitL_E=& _p[ _E-q_y(m)]_+- _1 (1-m)\\ &+ _b[ (1-m)-b]_+^2+ _2H(m). split (7) Here [a]+=max⁡(a,0)[a]_+= (a,0), b is the soft edit-fraction limit, and λp=λb=8 _p= _b=8 in the locked configuration. The counterfactual surrogate targets original-class probability below 0.49; actual hard success uses the classifier’s 0.5 decision rule. The factual quadratic term averages squared gates, whereas the counterfactual term squares the mean edited mass; the exemplar term is entropy regularization, not an elastic-net penalty. These losses are differentiable surrogates, not the final explanations. After optimization, the algorithm orders the unique undirected gates and evaluates hard masks in ranked order. It selects the smallest sufficient factual mask, the first prediction-changing deletion set, or the largest in-budget confidence-preserving exemplar deletion set. If no hard candidate satisfies its definition, the method returns a failure. This avoids the soft–hard mismatch that can arise when the optimized mask is thresholded. This deterministic feasibility audit extends CF-GNNExplainer’s best-candidate tracking to all three perspectives. These selections are extremal only within the ranked candidate sequence; neither the edge-mask scan nor the bounded fragment beam proves a global optimum of Eqs. (2)–(4). Algorithm 1 Multi-Perspective CF-GNN Explanation for Graph Classification 1: graph G=(V,E,X)G=(V,E,X) with M=|E|≥1M=|E|≥ 1; logit predictor fWf_W; view v∈F,C,Ev∈\F,C,E\; ρ; fraction b; absolute cap BmaxB_ ; β1,β2,λp,λb _1, _2, _p, _b; steps K; learning rate η; seed s 2: hard explanation GvG_v or an explicit failure 3: set predictor to evaluation mode; freeze every parameter in W 4: y←y⁡(G)y← y(G); τ←ρpy(G)τ←ρ p_y(G) 5: B←min⁡M,Bmax,max⁡(1,⌊bM⌋)B← \M,B_ , (1, bM )\ 6: create one logit zez_e for every unique undirected edge e∈Ee∈ E 7: initialize z from seed s 8: for t=1,…,Kt=1,…,K do 9: m←σ⁡(z)m←σ(z); mirror m onto both edge directions 10: qy←py(G⊙m)q_y← p_y(G m) using σ(fW(⋅))σ(f_W(·)) 11: update z by Adam to minimize Lv(qy,m)L_v(q_y,m) 12: end for 13: m←σ⁡(z)m←σ(z) after the final update 14: if v=Fv=F then 15: order edges by decreasing mem_e 16: ←C← retained-edge prefixes from smallest to largest 17: Gv←G_v← first H∈H with y⁡(H)=y(H)=y and py(H)≥τp_y(H)≥τ 18: else if v=Cv=C then 19: order edges by increasing mem_e 20: ←C← deletion prefixes of size 1,…,B1,…,B 21: Gv←G_v← first H∈H with y⁡(H)≠y(H)≠ y 22: else 23: order edges by increasing mem_e 24: ←C← deletion prefixes of size 1,…,B1,…,B 25: Gv←G_v← H∈H with the largest deletion prefix and y⁡(H)=y(H)=y and py(H)≥τp_y(H)≥τ 26: end if 27: if no candidate satisfies the hard constraints then 28: return Failure with the violated constraint 29: end if 30: return GvG_v, its hard mask, probabilities, edits, and provenance Candidates are symmetric deletions of existing undirected edges, satisfying structural validity by construction. Algorithm 1 specifies the proposed structural graph-classification extension of CF-GNNExplainer, not the fragment backend or the baselines. It optimizes only unique-edge gates and uses gradient descent through Adam; it does not optimize dense non-edge entries or update the predictor. Most importantly, the returned object is selected by an explicit hard feasibility scan rather than by thresholding a soft mask and assuming that its semantics survive discretization. A failure means that no searched candidate was feasible, not that no explanation exists in the full edit space. The structural backend masks edges but retains the full node set. The GCN also adds a self-loop at every node. Consequently, a one-edge factual mask can still use every node’s features through self-messages and should be interpreted as an edge-sufficiency result, not as a two-node induced subgraph or a chemically complete fragment. We therefore report retained edges together with isolated nodes and connected components, and reserve molecular-fragment language for the chemistry-aware backend below. IV-C Chemistry-aware fragment search Arbitrary deletion masks can leave atom sets that do not describe a useful molecule. For molecular graphs with SMILES provenance, we therefore implement a second search space. At each step the method cuts a non-ring single bond, retains a connected component, restores implicit hydrogens through canonicalization, and accepts the candidate only if RDKit sanitization succeeds. Duplicate canonical SMILES are removed. The search advances for at most three deletion steps. At each depth, at most 100 distinct candidates are retained, ordered by decreasing heavy-atom count and then canonical SMILES. Every retained fragment has at least two atoms; the unchanged query is not generated. Factual selection minimizes retained heavy atoms, breaking ties by fewer steps and greater query similarity. Counterfactual selection minimizes deletion steps, then removed heavy atoms, then prefers greater similarity. Exemplar selection maximizes fingerprint distance, breaking ties by more steps and fewer retained heavy atoms. Only counterfactual and exemplar candidates are restricted to removal of at most 35% of the query’s heavy atoms. Morgan fingerprints use radius two and the recorded RDKit default bit-vector settings. Sanitization checks representability and valence consistency [23]; it does not establish synthetic accessibility, likely occurrence, or a physically changed property. This backend adds chemical checks but is not a learned in-distribution generator. IV-D Predictive model The primary classifier is a three-layer GCN [24]. With A¯=A+I A=A+I and D¯ii=∑jA¯ij D_i= _j A_ij, a layer is Z(ℓ+1)=D¯−1/2A¯D¯−1/2H(ℓ)W(ℓ)+a(ℓ),Z^( +1)= D^-1/2 A D^-1/2H^( )W^( )+a^( ), (8) where a(ℓ)a^( ) is a bias. Each convolution is followed by BatchNorm, ReLU, and training-only dropout. Node representations are sum-pooled and passed to a linear binary-logit head; sigmoid obtains probabilities. Categorical node embeddings are summed column-wise; continuous node features use a linear encoder. The primary GCN consumes node features and topology, not edge labels or distance attributes. The GINE configuration used in validation sensitivity checks additionally encodes edge features; it is not a completed held-out architectural robustness study. V Experimental Protocol V-A Research questions and evaluation scope We organize the evaluation around five questions. RQ1 measures whether each hard output satisfies its factual, counterfactual, or exemplar definition. RQ2 evaluates compactness, faithfulness, and edit cost conditional on success while retaining failures in the coverage denominator. RQ3 tests whether the views select different structures on jointly successful cases. RQ4 measures repeatability across explainer seeds and sensitivity to the representation and edit budget. RQ5 tests how much counterfactual coverage changes when rankings learned by different objectives receive the same hard-prefix projection. This replay does not isolate every difference in optimization or training budget. V-B Datasets, features, and splits Table I summarizes the five evaluated datasets. MUTAG and Mutagenicity concern mutagenicity classification [25, 26]; AIDS provides a molecular activity classification benchmark through the IAM repository [27]. COX2_MD derives from COX2 activity data [28] and its distance-graph construction [29]: after removing explicit hydrogens, complete graphs carry interatomic-distance attributes and bond-type labels, including a no-bond category. Removing a COX2_MD graph edge is consequently not necessarily removing a chemical bond. We load all four from the official TU archives [30] and preserve source indices, categorical node and edge labels, and continuous attributes. Continuous attributes are standardized using training-fold statistics only. Stored edge labels and attributes are available to the GINE validation configuration but not consumed by the primary GCN. They are not explanation annotations; consequently, we do not report edge-mask accuracy on these datasets. For each TU dataset, one deterministic stratified ten-fold partition is created with master seed 20260911. In outer fold k, fold k is test, fold (k+1)mod10(k+1) 10 is validation, and the remaining folds are training. Every graph is therefore tested exactly once per predictor seed. Predictor metrics use all ten outer folds. Explanation generation is complete for all ten folds of MUTAG, Mutagenicity, and COX2_MD and for folds 0–7 of AIDS; the latter contains 1,600 distinct held-out graphs per predictor seed. The reporting configuration excludes the incomplete AIDS fold 8 as a whole and the unstarted fold 9, independent of their outcomes. BBBP uses its single locked scaffold test set. Thus every explanation table and plot uses only complete dataset–fold–predictor cells, with all attempted graphs and failed explanations retained inside those cells. BBBP is loaded from the processed MoleculeNet/PyG collection [1, 31]. We record the source row, canonical SMILES, Bemis–Murcko scaffold, label, and any additional loader rejection. Table I counts accepted graphs, not rejected or upstream-filtered source molecules. A deterministic stratified-group split groups equal scaffold strings, including empty strings for acyclic molecules. It selects test and validation folds closest to the target sizes and overall class balance, yielding approximately 80%, 10%, and 10% training, validation, and test partitions with zero scaffold overlap. Dataset manifests contain SHA-256 hashes of all available raw and processed files, feature schemas, graph identifiers, and split membership. TABLE I: Datasets and locked evaluation protocol. |E||E| counts each undirected edge once. N:L and E:L give node/edge categorical-label cardinalities; N:A and E:A give continuous-attribute dimensions. BBBP uses MoleculeNet atom/bond categorical features. GT denotes a documented edge-level explanation mask; none of these five datasets provides one. Features describe stored data, not all inputs used by GCN. AIDS explanation reporting uses folds 0–7; predictor reporting uses all ten folds. Dataset N Pos. |V||V| |E||E| Features Split GT MUTAG 188 0.66 17.9±4.617.9± 4.6 19.8±5.719.8± 5.7 N:L7; E:L4 Stratified 10-Fold no Mutagenicity 4337 0.45 30.3±20.130.3± 20.1 30.8±16.830.8± 16.8 N:L14; E:L3 Stratified 10-Fold no AIDS 2000 0.80 15.7±13.715.7± 13.7 16.2±15.016.2± 15.0 N:L38; N:A4; E:L3 Stratified 10-Fold no COX2_MD 303 0.49 26.3±2.526.3± 2.5 335.1±65.3335.1± 65.3 N:L7; E:L5; E:A1 Stratified 10-Fold no BBBP 2039 0.77 24.1±10.624.1± 10.6 26.0±11.726.0± 11.7 Atom/Bond Categorical Scaffold 80/10/10 no V-C Predictors and eligibility The frozen predictor is a three-layer GCN, with 64 hidden units per layer, BatchNorm, ReLU, dropout 0.2, sum pooling, and a binary logit head. Training uses Adam, class-weighted binary cross-entropy with the positive weight estimated from the training partition, learning rate 10−310^-3, weight decay 10−410^-4, batch size 64, at most 200 epochs, and early stopping after 30 epochs without improvement in validation log loss. Model seeds are 11, 29, and 47. We report test AUROC, AUPRC, balanced accuracy, F1, Brier score, expected calibration error, log loss, and the selected epoch before reporting any explanation result. All predictions in the locked explanation population are eligible for model-faithfulness analysis because the target is the frozen model’s prediction. The displayed dataset-level dagger is assigned when the mean balanced accuracy is below 0.55 or mean AUROC below 0.60. It cautions against application-utility interpretation; explanation eligibility is not conditioned on predictive correctness. V-D Hyperparameter selection and budgets No test graph participates in selection. We use at most 24 deterministically selected, class-balanced validation graphs per fold with predictor seed 11 and explainer seed 101 to evaluate the complete grid β1∈0.1,0.001,0.0001 _1∈\0.1,0.001,0.0001\ and β2∈0.1,0.01,0.001 _2∈\0.1,0.01,0.001\ for each perspective. For factual and counterfactual views, β1 _1 and β2 _2 both regularize retained mass for factuals and edited mass for counterfactuals; for exemplars they control the edit reward and binarization penalty. One global configuration per perspective is selected lexicographically by: (i) maximum hard-constraint success, (i) minimum retained fraction for factuals, minimum edit count for counterfactuals, or maximum feasible edit fraction for exemplars, (i) maximum retained query-class confidence, and (iv) minimum runtime. Dataset-level validation means are macro-averaged so Mutagenicity does not dominate smaller domains. We additionally evaluate edit-fraction budgets 0.05, 0.10, 0.20, 0.30, and 0.35. The default hard budget is B=min⁡|E|,8,max⁡(1,⌊0.35|E|⌋)B= \|E|,8, (1, 0.35|E| )\ and ρ=0.8ρ=0.8. The minimum-one clamp applies to non-empty structural graphs, relaxing the fractional limit when a graph is too small to admit one deletion. Explainer seeds are 101, 211, and 307. The MPGX-CF gate optimizer runs for 200 steps with Adam at learning rate 0.03. CF2 follows the official 500-step, 0.05 learning-rate graph configuration. PGExplainer is trained within each training partition for 20 epochs at learning rate 0.003; this training time is excluded from per-query explanation timings. V-E Comparison methods All compatible methods use the same predictor checkpoints, graph identifiers, splits, and output-feasibility criteria, but their native discretization rules differ. GNNExplainer and fold-trained PGExplainer are factual baselines [3, 4]. A graph-classification adaptation of CF-GNNExplainer is evaluated only for counterfactual search [5]. It uses the published negative-NLL prediction term while the thresholded label remains unchanged, plus β=0.5β=0.5 times the differentiable deletion distance. Following the authors’ released run commands, its symmetric gates are optimized for 500 SGD steps at learning rate 0.1 with Nesterov momentum 0.9; the best hard, in-budget flip is retained. CF2 uses the authors’ joint L1+λ⁡[αLf+(1−α)Lc]L_1+λ[α L_f+(1-α)L_c] objective, initialization, 0.5 mask threshold, and published graph defaults (λ=1000λ=1000, α=0.6α=0.6, γ=0.5γ=0.5); the same learned mask is evaluated as a factual subgraph and its complement for necessity [8]. A one-edge-deletion greedy ranking is the exemplar control. Matched random rankings are evaluated for all three perspectives. To assess this discretization confound, we replay stored counterfactual rankings through one identical hard-prefix search on a deterministic, class-balanced subset of 24 held-out graphs per dataset (12 per original predicted class), with all three GCN seeds and explainer seed 101. The subset contains only graph IDs with complete records for all compared rankings. We also evaluate the proposed soft mask at a fixed 0.5 threshold, and test both a node-feature-only zero-edge input and an induced-node version of each selected factual explanation. These are exploratory controls, not a replacement for the locked full-population experiment. For BBBP, we additionally report a separate sanitized fragment-deletion search. RDKit validity establishes graph representability, not synthesizability, and these chemistry-aware results are not pooled with the structural edge-mask comparison. V-F Outcomes and statistical analysis Semantic success is the hard label/confidence predicate of the requested view. Factual quality uses retained-edge fraction and comprehensiveness, py(G)−py(G∖EF)p_y(G)-p_y(G E_F), where the complement retains all query nodes. Counterfactual quality uses deleted-edge count and original-class probability drop, py(G)−py(H)p_y(G)-p_y(H). Exemplar quality uses deleted-edge fraction and retained-confidence ratio, py(H)/py(G)p_y(H)/p_y(G). Structural edits are deletion-only; no edge-addition outcome or geometric boundary-distance estimate is inferred. Components and isolated nodes are counted on the full retained node set without predictor-added self-loops. Probability-shift diagnostics use |p⁡(H)−p⁡(G)||p(H)-p(G)|. Centrality diagnostics are root-mean-square nodewise changes in NetworkX-normalized betweenness and unnormalized harmonic centrality, not chemical importance scores. Stability averages pairwise Jaccard similarity across successful explainer repetitions for the same graph–model pair, using retained edges for factuals and deleted edges for other views. Units with fewer than two successful repetitions supply no stability value. Cross-perspective Jaccard overlap compares factual retained edges with counterfactual or exemplar deleted edges on complete successful triplets. Runtime is wall time per query and view, including failures, and is distinct from the sum across views. Timing depends on implementation and hardware; it is not a universal complexity comparison. Failed searches remain in the denominator and retain a machine-readable reason. Explainer repetitions are averaged within each graph–model pair. Main-text continuous outcomes are reported as mean ± standard deviation over these analysis units, while success is reported as a percentage of all attempts, including failed searches. Descriptive intervals in the supplementary metric records use 10,000 hierarchical bootstrap resamples over fold, model seed, and graph as applicable; new timing intervals use the same hierarchy. Paired effect intervals instead resample unique held-out graphs after averaging their repeated predictor-seed differences. Method tests use these paired graph-level differences on shared examples, Wilcoxon signed-rank tests, and Holm correction within each dataset–perspective–metric family. Datasets are separate domains; no micro-average across datasets is used. Shared fitted models induce dependence among held-out graphs, so intervals and paired tests are descriptive of the evaluated population, not guarantees of generalization to new training samples. V-G Qualitative selection and reproducibility The overview prefers complete successful MPGX-CF triplets selected by minimum distance to dataset medians of mean edit count and mean retained query-class probability. If none exist, it retains cases with the most successful views before applying the same rule. This deterministic illustration rule is not a claim of prospective preregistration. The appendices provide enlarged views and a failure audit. TU panels display documented categorical labels and are not described as sanitized molecules. Every panel is regenerated from its graph ID and hard mask and records fold, predictor seed, explainer seed, probabilities, edit count, and failure status. The reported experiment is staged as predictor training, validation-only selection, locked test explanation, and artifact generation. Runs are resumable by dataset, fold, architecture, predictor seed, explainer seed, graph, method, perspective, budget, and protocol hash. The local reproducibility package records the resolved configuration, environment, source commit (or SHA-256 source hashes when commit metadata are unavailable), checkpoints, histories, split manifests, failures, and long-form CSV/Parquet output. All dependencies are installed in project-local UV environments. The arXiv source archive contains the manuscript, rendered tables and figures, a claim audit, supplementary metric summaries, and compact provenance CSVs, but not the large raw experiment directory; a public accession for the latter has not been assigned. VI Results and Discussion VI-A Predictor performance and calibration Predictive results establish the context for the explanation study. A dagger marks a dataset-level mean that triggers the weak-predictor rule; its explanations characterize the frozen model, not established application utility. TABLE I: Primary GCN out-of-fold or scaffold-held-out predictive performance. Values are mean ± standard deviation over outer-fold/model-seed runs. Dataset Architecture Balanced accuracy AUROC AUPRC F1 AIDS GCN 0.971±0.0140.971± 0.014 0.985±0.0150.985± 0.015 0.991±0.0120.991± 0.012 0.989±0.0050.989± 0.005 BBBP GCN 0.698±0.0140.698± 0.014 0.797±0.0100.797± 0.010 0.942±0.0080.942± 0.008 0.907±0.0080.907± 0.008 COX2_MD† GCN 0.519±0.0690.519± 0.069 0.561±0.1330.561± 0.133 0.586±0.0990.586± 0.099 0.522±0.1810.522± 0.181 MUTAG GCN 0.688±0.1600.688± 0.160 0.836±0.1020.836± 0.102 0.929±0.0430.929± 0.043 0.835±0.0620.835± 0.062 Mutagenicity GCN 0.781±0.0240.781± 0.024 0.856±0.0220.856± 0.022 0.819±0.0290.819± 0.029 0.757±0.0270.757± 0.027 TABLE I: Predictive calibration and selected training epoch; mean ± standard deviation over the same outer-fold/model-seed runs. Dataset Architecture Brier ECE Log loss Best epoch AIDS GCN 0.015±0.0060.015± 0.006 0.023±0.0070.023± 0.007 0.077±0.0500.077± 0.050 37.8 BBBP GCN 0.122±0.0050.122± 0.005 0.094±0.0100.094± 0.010 0.466±0.0350.466± 0.035 10.3 COX2_MD GCN 0.285±0.0650.285± 0.065 0.204±0.1270.204± 0.127 0.865±0.3820.865± 0.382 6.9 MUTAG GCN 0.162±0.0450.162± 0.045 0.204±0.0460.204± 0.046 0.494±0.1140.494± 0.114 51.2 Mutagenicity GCN 0.153±0.0120.153± 0.012 0.052±0.0120.052± 0.012 0.494±0.0530.494± 0.053 86.8 Held-out GCN AUROC ranged from 0.561±0.1330.561± 0.133 on COX2_MD to 0.985±0.0150.985± 0.015 on AIDS. The prespecified weak-predictor rule flagged COX2_MD; its explanation results characterize that frozen model rather than application utility. VI-B Semantic coverage and conditional quality (RQ1–RQ2) Table IV reports all three proposed explanations for every dataset. Success uses all attempted graph–model–explainer units as its denominator; the remaining quantities are conditional on success and are therefore never interpreted without coverage. TABLE IV: Primary GCN multi-perspective explanations on all five datasets. Success is the percentage of all attempts; continuous outcomes are mean ± standard deviation after averaging explainer repetitions within each graph–model pair. Comp. denotes comprehensiveness and Δpy p_y is the drop in the original-class probability. Factual Counterfactual Exemplar Dataset Success Retained Comp. Success Edits Δpy p_y Success Edited Retained pyp_y ratio MUTAG 100.0% 0.105±0.1310.105± 0.131 0.120±0.2200.120± 0.220 67.6% 2.59±1.572.59± 1.57 0.436±0.1920.436± 0.192 98.9% 0.279±0.0480.279± 0.048 1.041±0.1611.041± 0.161 Mutagenicity 100.0% 0.155±0.2030.155± 0.203 0.142±0.2410.142± 0.241 63.7% 2.85±1.912.85± 1.91 0.315±0.1550.315± 0.155 99.0% 0.246±0.0680.246± 0.068 1.046±0.1291.046± 0.129 AIDS 100.0% 0.103±0.0790.103± 0.079 0.019±0.1130.019± 0.113 4.8% 3.48±2.103.48± 2.10 0.419±0.1650.419± 0.165 100.0% 0.285±0.0590.285± 0.059 1.002±0.0401.002± 0.040 COX2_MD 100.0% 0.092±0.2180.092± 0.218 0.048±0.1730.048± 0.173 5.5% 4.68±2.334.68± 2.33 0.099±0.1140.099± 0.114 100.0% 0.025±0.0050.025± 0.005 0.998±0.0280.998± 0.028 BBBP 100.0% 0.086±0.0930.086± 0.093 0.090±0.1630.090± 0.163 54.6% 3.65±2.243.65± 2.24 0.382±0.1460.382± 0.146 99.7% 0.277±0.0520.277± 0.052 1.008±0.0941.008± 0.094 Across datasets, hard factual coverage was 100.0% and successful factuals retained 0.09–0.16 of input edges. Counterfactual flip coverage was 4.8–67.6%; successful flips removed 2.59–4.68 edges on average, depending on the dataset. Exemplar preservation coverage was 98.9–100.0%; successful exemplars edited 0.02–0.29 of edges while retaining 1.00–1.05 of the query-class confidence. These conditional ranges are interpreted only together with their coverage denominators. Factual coverage must be read in light of the candidate family: its full-graph prefix reproduces the query prediction. Complete coverage is therefore expected, not evidence that the learned ranking is uniquely informative. Compactness and comprehensiveness characterize the smaller returned mask. Counterfactual and exemplar searches instead require a positive, bounded edit and may have no successful candidate. High exemplar coverage consequently indicates prediction-preserving edits in this space, not general model robustness or stable attribution. Interpreting factual sparsity. The factual projection deliberately scans retained-edge prefixes from one edge upward and accepts the first mask that preserves the label and at least 0.8py(G)0.8p_y(G). Validation selection then prefers the smallest retained fraction among successful configurations. Very sparse factual masks are therefore an expected consequence of the stated edge-sufficiency objective when node features and self-loops dominate; they are not, by themselves, evidence of a small connected mechanism. Table XIII reports isolated nodes and components, including nodes absent from retained-edge incidence. The enlarged panels keep these nodes visible in grey rather than silently displaying an induced subgraph. The method-appropriate factual, counterfactual, and exemplar native-output comparisons are reported separately in Tables V– VII; an empty method–view combination is never manufactured by reinterpreting another algorithm’s output. Because the counterfactual discretizers differ, these native-output comparisons are descriptive, not an attribution of a coverage difference to the learned loss. TABLE V: Factual native-output comparisons. Success includes failures; quality is conditional on success and uses graph–model means ± standard deviation after averaging explainer repetitions. Dataset Method Success Retained Comp. MUTAG MPGX-CF 100.0% 0.105±0.1310.105± 0.131 0.120±0.2200.120± 0.220 MUTAG GNNExplainer 100.0% 0.160±0.2100.160± 0.210 0.103±0.2130.103± 0.213 MUTAG PGExplainer 100.0% 0.219±0.2800.219± 0.280 0.112±0.2430.112± 0.243 MUTAG CF2 reimplementation 96.8% 0.445±0.1540.445± 0.154 0.519±0.3550.519± 0.355 MUTAG Random 100.0% 0.183±0.2220.183± 0.222 0.114±0.2200.114± 0.220 Mutagenicity MPGX-CF 100.0% 0.155±0.2030.155± 0.203 0.142±0.2410.142± 0.241 Mutagenicity GNNExplainer 100.0% 0.209±0.2620.209± 0.262 0.118±0.2260.118± 0.226 Mutagenicity PGExplainer 100.0% 0.266±0.3310.266± 0.331 0.140±0.2470.140± 0.247 Mutagenicity CF2 reimplementation 96.8% 0.579±0.1530.579± 0.153 0.417±0.2830.417± 0.283 Mutagenicity Random 100.0% 0.229±0.2810.229± 0.281 0.103±0.2140.103± 0.214 AIDS MPGX-CF 100.0% 0.103±0.0790.103± 0.079 0.019±0.1130.019± 0.113 AIDS GNNExplainer 100.0% 0.112±0.1120.112± 0.112 0.018±0.1090.018± 0.109 AIDS PGExplainer 100.0% 0.111±0.1150.111± 0.115 0.014±0.1030.014± 0.103 AIDS CF2 reimplementation 61.7% 0.405±0.1860.405± 0.186 0.094±0.2120.094± 0.212 AIDS Random 100.0% 0.113±0.1180.113± 0.118 0.017±0.1060.017± 0.106 COX2_MD MPGX-CF 100.0% 0.092±0.2180.092± 0.218 0.048±0.1730.048± 0.173 COX2_MD GNNExplainer 100.0% 0.115±0.2480.115± 0.248 0.057±0.1780.057± 0.178 COX2_MD PGExplainer 100.0% 0.177±0.3220.177± 0.322 0.072±0.2060.072± 0.206 COX2_MD CF2 reimplementation 56.5% 0.112±0.1510.112± 0.151 0.139±0.2640.139± 0.264 COX2_MD Random 100.0% 0.085±0.1920.085± 0.192 0.031±0.1300.031± 0.130 BBBP MPGX-CF 100.0% 0.086±0.0930.086± 0.093 0.090±0.1630.090± 0.163 BBBP GNNExplainer 100.0% 0.145±0.1690.145± 0.169 0.071±0.1280.071± 0.128 BBBP PGExplainer 100.0% 0.186±0.2480.186± 0.248 0.072±0.1410.072± 0.141 BBBP CF2 reimplementation 95.3% 0.542±0.1220.542± 0.122 0.521±0.2690.521± 0.269 BBBP Random 100.0% 0.182±0.2100.182± 0.210 0.067±0.1320.067± 0.132 TABLE VI: Counterfactual native-output comparisons under the locked budget. Success includes failures; quality is conditional on success. Native discretization and optimization procedures differ across methods. Dataset Method Success Edits Δpy p_y MUTAG MPGX-CF 67.6% 2.594±1.5692.594± 1.569 0.436±0.1920.436± 0.192 MUTAG CF-GNNExplainer 3.7% 1.190±0.4021.190± 0.402 0.398±0.1180.398± 0.118 MUTAG CF2 reimplementation 7.2% 5.854±1.6635.854± 1.663 0.756±0.2020.756± 0.202 MUTAG Random 31.3% 3.452±1.9103.452± 1.910 0.396±0.1870.396± 0.187 Mutagenicity MPGX-CF 63.7% 2.851±1.9122.851± 1.912 0.315±0.1550.315± 0.155 Mutagenicity CF-GNNExplainer 2.2% 1.286±0.5901.286± 0.590 0.335±0.1270.335± 0.127 Mutagenicity CF2 reimplementation 1.3% 5.281±2.0985.281± 2.098 0.321±0.2090.321± 0.209 Mutagenicity Random 22.5% 3.568±1.9373.568± 1.937 0.231±0.1400.231± 0.140 AIDS MPGX-CF 4.8% 3.475±2.1003.475± 2.100 0.419±0.1650.419± 0.165 AIDS CF-GNNExplainer 0.1% 1.000±0.0001.000± 0.000 0.380±0.0900.380± 0.090 AIDS CF2 reimplementation 0.2% 4.300±1.9254.300± 1.925 0.408±0.2680.408± 0.268 AIDS Random 1.2% 3.628±2.1463.628± 2.146 0.330±0.1730.330± 0.173 COX2_MD MPGX-CF 5.5% 4.679±2.3284.679± 2.328 0.099±0.1140.099± 0.114 COX2_MD CF-GNNExplainer 0.0% – – COX2_MD CF2 reimplementation 0.3% 6.667±0.4716.667± 0.471 0.041±0.0520.041± 0.052 COX2_MD Random 1.0% 5.678±1.9345.678± 1.934 0.060±0.0590.060± 0.059 BBBP MPGX-CF 54.6% 3.647±2.2443.647± 2.244 0.382±0.1460.382± 0.146 BBBP CF-GNNExplainer 0.2% 1.000±0.0001.000± 0.000 0.369±0.0000.369± 0.000 BBBP CF2 reimplementation 0.5% 6.250±1.2556.250± 1.255 0.466±0.2420.466± 0.242 BBBP Random 10.6% 3.951±2.0543.951± 2.054 0.238±0.1310.238± 0.131 TABLE VII: Exemplar comparisons. Coverage requires a non-trivial prediction-preserving edit; conditional quality uses the same aggregation as Table IV. Dataset Method Success Edited Retained pyp_y ratio MUTAG MPGX-CF 98.9% 0.279±0.0480.279± 0.048 1.041±0.1611.041± 0.161 MUTAG Greedy control 100.0% 0.305±0.0260.305± 0.026 1.169±0.2321.169± 0.232 MUTAG Random 94.6% 0.268±0.0600.268± 0.060 1.023±0.1451.023± 0.145 Mutagenicity MPGX-CF 99.0% 0.246±0.0680.246± 0.068 1.046±0.1291.046± 0.129 Mutagenicity Greedy control 99.9% 0.257±0.0690.257± 0.069 1.191±0.2111.191± 0.211 Mutagenicity Random 96.5% 0.231±0.0740.231± 0.074 1.027±0.1271.027± 0.127 AIDS MPGX-CF 100.0% 0.285±0.0590.285± 0.059 1.002±0.0401.002± 0.040 AIDS Greedy control 100.0% 0.287±0.0580.287± 0.058 1.021±0.0651.021± 0.065 AIDS Random 99.7% 0.285±0.0590.285± 0.059 1.003±0.0401.003± 0.040 COX2_MD MPGX-CF 100.0% 0.025±0.0050.025± 0.005 0.998±0.0280.998± 0.028 COX2_MD Greedy control 100.0% 0.025±0.0050.025± 0.005 1.018±0.0591.018± 0.059 COX2_MD Random 100.0% 0.025±0.0050.025± 0.005 0.998±0.0270.998± 0.027 BBBP MPGX-CF 99.7% 0.277±0.0520.277± 0.052 1.008±0.0941.008± 0.094 BBBP Greedy control 100.0% 0.286±0.0520.286± 0.052 1.137±0.1861.137± 0.186 BBBP Random 98.9% 0.268±0.0590.268± 0.059 0.998±0.0840.998± 0.084 The exemplar control is particularly informative: a simple greedy deletion ranking matches or exceeds the proposed method’s preservation coverage and edited fraction on these datasets (Table VII). The three-view contribution is therefore a common semantic and audit protocol, not evidence that the learned exemplar ranking is superior to this control. VI-C Complementarity and repeatability (RQ3–RQ4) Table XV compares factual retained edges with the deletions returned by the other views on complete successful triplets. Mean factual–counterfactual overlap exceeds factual–exemplar overlap in each dataset. This is consistent with distinct roles for decision-changing and prediction-preserving deletions, but their differing sizes and the successful-triplet restriction affect Jaccard values. It does not establish independent explanations or distinct biochemical mechanisms. Prediction preservation is not mask repeatability. Mean seed-level Jaccard stability ranges across datasets from 0.06–0.42 for factual retained sets, 0.31–0.94 for counterfactual deletions, and 0.02–0.24 for exemplar deletions. These stability values are conditional on having at least two successful repetitions, not on all attempted searches. In particular, the low factual and exemplar repeatability means that multiple edge sets can satisfy similar prediction constraints. These outputs are alternative audits of one classifier, not uniquely recovered structural truth. Conditional centrality changes and per-view runtime are reported in the appendix; runtime includes failed searches and should be interpreted in the recorded execution environment. VI-D Exploratory projection and representation controls (RQ4–RQ5) Table IX re-evaluates stored counterfactual rankings on the same 24 held-out graph IDs per dataset with three predictor seeds (7272 graph–model units per dataset). Every method receives the same ranked deletion-prefix scan and the same eight-edge and 35%-of-edges caps. In this controlled subset, the same learned scores can yield substantially different coverage when their hardening procedure changes. On the common-projection BBBP subset, CF2 coverage changes from 2.8% to 63.9%, compared with 65.3% for the proposed ranking. CF-GNNExplainer changes from 1.4% to 54.2%. On AIDS, CF2 and the proposed ranking both reach 6.9%. The native-output gap therefore cannot be assigned solely to the learned objective. The proposed 0.5-threshold ablation and detailed protocol are reported in Appendix C. The low counterfactual coverage on AIDS and COX2_MD is not a universal failure rate of graph counterfactual search: it is tied to this frozen predictor and deletion budget. On the same selected units, increasing the absolute cap from 8 to 32 while retaining the fractional cap changes COX2_MD coverage from 9.7% to 23.6%; AIDS remains at 6.9%. This is a budget sensitivity result, not a redefinition of the primary outcome. COX2_MD predictor AUROC is also weak, so these explanations are useful for auditing that model, not for chemical conclusions. The factual null further distinguishes edge sufficiency from a molecular fragment. The zero-edge input, which keeps node features and predictor self-loops, satisfies the factual label/confidence predicate for 48.6%–91.7% of selected graph–model units across datasets. With only nodes incident to the selected factual edges retained, sufficiency is 37.5%–61.1%. The latter intervention changes the model input and does not itself produce a validated molecular fragment. These observations support the full-node interpretation of the mask backend. Full values appear in Appendix C. Figure 1 presents semantic coverage before conditional quality. Figure 4 then reports the structural and probabilistic outcomes of successful explanations with standard-deviation error bars. This order prevents a method with low coverage from appearing strong merely because its unsuccessful cases were removed. Fig. 1: Hard-constraint coverage for compatible methods on all five datasets. Bars are means and error bars are standard deviations over graph–model analysis units after averaging explainer repetitions. The appendix statistical tables contain conditional compactness, stability, structural diagnostics, runtime, cross-perspective overlap, paired effect estimates, and Holm-adjusted p values. No analogous attribution accuracy is reported for TU edge labels because they encode input categories, not explanation annotations. VI-E Qualitative successes and failures Figure 2 presents all three explanation views for all five datasets. It is generated by the deterministic median rule. Red dashed edges are excluded from the returned graph and retained edges are green. When no complete successful triplet exists, the affected panel is marked as a failure rather than removing the dataset row. The appendices also provide one full-width strip per dataset in the visual style of the primary BBBP case study, with the same locked cases and provenance. Fig. 2: Input, factual, counterfactual, and exemplar panels for MUTAG, Mutagenicity, AIDS, COX2_MD, and BBBP. A common graph and fixed layout are used within each row; panels print class-1 probability, label, and deleted unique-edge count. TU graphs are categorical renderings, not reconstructed or sanitized molecules. Unavailable rows remain visible rather than being silently omitted; the primary chemistry-aware BBBP case is shown separately in Fig. 3. Selections are MUTAG: mutag-000074, fold 5, model seed 29, explainer seed 101; Mutagenicity: mutagenicity-003801, fold 7, model seed 47, explainer seed 211; AIDS: aids-001512, fold 0, model seed 11, explainer seed 101; COX2_MD: cox2_md-000133, fold 3, model seed 29, explainer seed 307; BBBP: bbbp-001663, fold 0, model seed 47, explainer seed 307. Figure 3 reports a primary-test BBBP chemistry-aware case. Unlike the edge-mask panels above, every returned object is an RDKit-sanitized retained fragment. Its factual criterion minimizes retained heavy atoms rather than retained edges. The case illustrates the three semantic outcomes; population-level coverage and quality remain in Table VIII. Sanitation makes this an analysis of representable molecular variants, not experimental evidence that the predicted property changes in a synthesized compound. Fig. 3: Primary-test chemistry-aware BBBP fragment explanations. The factual and exemplar fragments preserve the query decision, whereas the counterfactual fragment reverses it. Probabilities printed in the panels are the frozen GCN’s class-1 probabilities. Graph bbbp-001204, scaffold-held-out BBBP test set, GCN predictor seed 11 and fragment-search seed 0; selected by median counterfactual Tanimoto similarity among complete successful fragment triplets in the primary run. Fig. 4: Conditional quality of successful factual, counterfactual, and exemplar explanations. Structural size and prediction behavior are shown together because neither alone establishes semantic validity. Diamonds share the fractional axis except in the counterfactual panel, where they use the right-hand probability-drop axis. VII Limitations and Responsible Use Model and representation dependence. MPGE explains a trained classifier, not a causal chemical mechanism. A factual mask can preserve a prediction through node features and self-messages without identifying a compact molecular fragment. COX2_MD also uses complete distance graphs whose edge attributes are not consumed by the primary GCN. Weak predictive performance, calibration errors, and altered-input distributions limit chemical interpretation [2, 11]. Search and evaluation scope. Ranked prefixes and fragment beams explore restricted deletion families; success verifies a returned candidate, whereas failure is not proof of global infeasibility. Seed-dependent masks can satisfy similar constraints, and prediction preservation does not certify robust attribution [20, 21]. Quality comparisons are conditional on success, native methods have different discretizers and optimization budgets, and the exploratory controls cover a selected subset. The primary study has three GCN seeds, eight completed AIDS explanation folds, and one BBBP scaffold split. GINE validation checks do not establish held-out architectural robustness or performance on new training samples. Chemical and human validity. RDKit sanitation establishes representability, not synthesizability, stability, safety, or experimentally changed activity [23]. The fragment backend cannot search atom or bond additions and does not enforce a learned molecular distribution. Expert-chemist evaluation and experimental property validation remain necessary before claiming medicinal-chemistry utility [2]. The figures support model inspection and hypothesis generation, not chemical intervention recommendations. Independent reproducibility. The upload provides generated artifacts, metric summaries, and compact claim and case provenance. The absence of a public row-level record and checkpoint accession still limits independent reruns; local completeness should not be equated with a public reproducibility release. VIII Conclusion MPGE connects three questions about one molecular graph prediction: what information supports it, what bounded deletion changes it, and what non-trivial deletion preserves it. View-specific objectives and hard-output checks make these questions operational in a graph-classification extension of CF-GNNExplainer, with a separate sanitized-fragment instantiation for BBBP. The evaluated GCNs yielded 4.8–67.6% bounded counterfactual coverage and 98.9–100.0% non-trivial exemplar coverage across the five datasets, illustrating why semantic coverage and conditional quality must be reported separately. Factual compactness is informative only under the documented full-node mask semantics, and high preservation coverage does not imply stable masks. The exploratory controls further show that discretization, retained features, and edit budgets materially shape the evidence. The resulting framework supports transparent multi-perspective model auditing; establishing broader architectural, chemical, or human utility remains a separate validation task. Appendix A Chemistry-Aware BBBP Fragment Analysis The fragment backend cuts only non-ring single bonds, canonicalizes each retained component, restores implicit hydrogens, and accepts only candidates that pass RDKit sanitization. Its search space and edit unit differ from the edge-mask backend, so these results are reported separately rather than pooled with the structural comparison. Failures remain in the coverage denominator; conditional rows summarize only returned sanitized molecules. Tanimoto similarity uses radius-two Morgan fingerprints, and retained confidence is relative to the unmodified query. The primary-test qualitative fragment triplet appears once, in Fig. 3. It is kept separate from the structural edge-mask strips because sanitization, fragment retention, and atom removal define a different intervention space. TABLE VIII: GCN chemistry-aware fragment search on the scaffold-held-out BBBP test set. Success is percentage-only over all attempts, including failures; other values are conditional means ± standard deviation over graph–model units. Fragment factual size is retained heavy atoms, not retained edges. View Success Steps Atoms removed Tanimoto Retained pyp_y ratio Candidates Factual 97.7% 1.71±0.701.71± 0.70 19.91±8.7919.91± 8.79 0.10±0.110.10± 0.11 0.92±0.120.92± 0.12 56.98±59.7556.98± 59.75 Counterfactual 27.4% 1.40±0.611.40± 0.61 3.22±2.663.22± 2.66 0.62±0.140.62± 0.14 0.56±0.150.56± 0.15 73.96±66.9773.96± 66.97 Exemplar 96.4% 2.19±0.782.19± 0.78 6.63±2.606.63± 2.60 0.35±0.120.35± 0.12 1.00±0.141.00± 0.14 57.54±59.9557.54± 59.95 Appendix B Validation-Only Selection The complete 3×33× 3 perspective-specific regularization grid is evaluated independently for each perspective on deterministic class-balanced validation subsets. The GCN grid sets the primary test configuration; GINE validation results are displayed as an architectural validation sensitivity check, not a completed GINE held-out explanation comparison. The factual and counterfactual losses use size/quadratic penalties; the exemplar uses an edit reward and entropy penalty. Test graphs never participate in selection. Fig. 5: Macro-averaged hard-constraint success over the complete GCN and GINE validation grids. Failed searches remain in every denominator. Fig. 6: Validation coverage as the permitted edit fraction changes. Error bars show variation across dataset-level means. Appendix C Exploratory Held-Out Controls The control population contains 24 graph IDs per dataset, selected with fixed seed 20261007 by predicted-class balance from complete counterfactual-ranking records. All selected graphs belong to their outer-fold test set or BBBP scaffold test set. Three frozen GCN checkpoints and one explainer seed (101) yield 72 graph–model units per dataset. The common projection scans deletion prefixes of stored edge rankings and evaluates every hard candidate with the same checkpoint. This controls the discretization rule but does not equate optimization budgets, objectives, or training procedures. It is exploratory and is not used to revise validation-selected hyperparameters. TABLE IX: Exploratory counterfactual control on a class-balanced held-out subset: native hard output versus identical ranked-prefix projection of each method’s stored scores. Percentages include failures; 72 graph–model units represent 24 graph IDs and three predictor seeds per dataset. Dataset Ranking Units Native Common projection MUTAG MPGX-CF 72 54.2% 54.2% MUTAG CF-GNNExplainer 72 5.6% 41.7% MUTAG CF2 72 4.2% 51.4% MUTAG Random 72 26.4% 26.4% Mutagenicity MPGX-CF 72 63.9% 63.9% Mutagenicity CF-GNNExplainer 72 1.4% 51.4% Mutagenicity CF2 72 2.8% 59.7% Mutagenicity Random 72 19.4% 19.4% AIDS MPGX-CF 72 6.9% 6.9% AIDS CF-GNNExplainer 72 0.0% 5.6% AIDS CF2 72 0.0% 6.9% AIDS Random 72 1.4% 1.4% COX2_MD MPGX-CF 72 9.7% 9.7% COX2_MD CF-GNNExplainer 72 0.0% 8.3% COX2_MD CF2 72 1.4% 6.9% COX2_MD Random 72 1.4% 1.4% BBBP MPGX-CF 72 65.3% 65.3% BBBP CF-GNNExplainer 72 1.4% 54.2% BBBP CF2 72 2.8% 63.9% BBBP Random 72 11.1% 11.1% The proposed soft mask itself flips far less often if a fixed 0.5 threshold is used in place of ranked-prefix search (Table X). Thus the search/projection mechanism is a major part of the implemented method, not a negligible post-processing detail. TABLE X: Proposed counterfactual ranking: fixed threshold versus hard-prefix search on the same selected held-out units. Dataset Units Proposed 0.5 threshold Proposed ranked prefix MUTAG 72 16.7% 54.2% Mutagenicity 72 9.7% 63.9% AIDS 72 1.4% 6.9% COX2_MD 72 0.0% 9.7% BBBP 72 15.3% 65.3% TABLE XI: Factual null and induced-node controls. The no-edge input retains all query nodes, features, and predictor self-loops; the induced input retains only nodes incident to the proposed factual edges. Percentages use all 72 graph–model units per dataset. Dataset Graphs Units No-edge feasible Induced factual feasible MUTAG 24 72 73.6% 37.5% Mutagenicity 24 72 56.9% 55.6% AIDS 24 72 91.7% 61.1% COX2_MD 24 72 69.4% 56.9% BBBP 24 72 48.6% 38.9% The eight-edge hard cap allows a small edit fraction in dense COX2_MD graphs. Table XII keeps the same learned rankings and 35% relative budget while changing only the absolute cap from 8 to 32. Coverage increases on COX2_MD, whereas AIDS remains unchanged. The latter observation alone does not identify which model or search constraint prevents additional flips. This is a sensitivity analysis, not a redefinition of the locked primary outcome. The full row-level records, selected graph IDs, and selection rule are in reviewer_controls. TABLE XII: Exploratory hard-budget sensitivity under a common projection on the same class-balanced held-out subset. Dataset Ranking Units Cap 8 Cap 32 AIDS MPGX-CF 72 6.9% 6.9% AIDS CF-GNNExplainer 72 5.6% 5.6% AIDS CF2 72 6.9% 6.9% AIDS Random 72 1.4% 1.4% COX2_MD MPGX-CF 72 9.7% 23.6% COX2_MD CF-GNNExplainer 72 8.3% 22.2% COX2_MD CF2 72 6.9% 22.2% COX2_MD Random 72 1.4% 5.6% Appendix D Statistical Tables For each continuous metric, explainer repetitions are first averaged within a graph–model pair. Descriptive intervals use 10,000 hierarchical bootstrap resamples. For paired method effects, repeated predictor-seed differences are averaged within each held-out graph before a graph-level bootstrap and Wilcoxon signed-rank test. Thus a graph, not a graph–model pair, is the resampling and test unit; shared fitted models can still induce dependence. Holm adjustment is performed within each dataset–architecture–perspective–outcome family. The local analysis also records graph count, model-seed count, and failure denominator. Dataset domains are never micro-averaged. The machine-readable sources are model_summary.parquet, explanation_summary.parquet, and paired_method_comparisons.parquet. The upload’s manumetric_summary.csv includes estimates, descriptive intervals, observation counts, analysis-unit counts, and success conditioning. Hierarchical interval centers weight folds and predictor seeds equally, while descriptive means weight available graph–model units equally; these estimands need not coincide when fold sizes or successful populations differ. TABLE XIII: Proposed GCN per-view runtime and connectivity. Runtime is unconditional wall time in seconds, including failed searches; components and isolated nodes are conditional on hard success, with predictor self-loops excluded. Values are mean ± standard deviation after averaging explainer seeds within graph–model units. Units counts timing units, not independent training replications. Runtime intervals are descriptive hierarchical 95% bootstrap intervals; hardware and implementation limit timing comparisons. Dataset View Runtime (s) Runtime 95% CI Components Isolated nodes Units MUTAG Factual 0.54±0.040.54± 0.04 [0.53,0.56][0.53,0.56] 15.99±4.9515.99± 4.95 14.64±5.1214.64± 5.12 564 MUTAG Counterfactual 0.55±0.040.55± 0.04 [0.54,0.55][0.54,0.55] 2.00±1.002.00± 1.00 0.52±0.790.52± 0.79 564 MUTAG Exemplar 0.57±0.040.57± 0.04 [0.56,0.58][0.56,0.58] 4.02±1.084.02± 1.08 1.46±0.781.46± 0.78 564 Mutagenicity Factual 0.58±0.220.58± 0.22 [0.51,0.68][0.51,0.68] 26.02±19.5626.02± 19.56 23.52±19.5823.52± 19.58 13011 Mutagenicity Counterfactual 0.56±0.160.56± 0.16 [0.51,0.64][0.51,0.64] 3.57±6.143.57± 6.14 1.79±6.041.79± 6.04 13011 Mutagenicity Exemplar 0.58±0.160.58± 0.16 [0.52,0.65][0.52,0.65] 6.79±9.106.79± 9.10 3.92±8.993.92± 8.99 13011 AIDS Factual 0.55±0.170.55± 0.17 [0.49,0.64][0.49,0.64] 14.45±13.6914.45± 13.69 13.35±13.7113.35± 13.71 4800 AIDS Counterfactual 0.54±0.150.54± 0.15 [0.49,0.62][0.49,0.62] 3.31±1.643.31± 1.64 1.08±1.071.08± 1.07 4800 AIDS Exemplar 0.56±0.140.56± 0.14 [0.50,0.63][0.50,0.63] 3.79±1.543.79± 1.54 1.51±0.991.51± 0.99 4800 COX2_MD Factual 0.63±0.160.63± 0.16 [0.58,0.69][0.58,0.69] 19.78±9.6419.78± 9.64 18.61±9.8218.61± 9.82 909 COX2_MD Counterfactual 0.59±0.070.59± 0.07 [0.56,0.63][0.56,0.63] 1.00±0.001.00± 0.00 0.00±0.000.00± 0.00 909 COX2_MD Exemplar 0.61±0.070.61± 0.07 [0.58,0.65][0.58,0.65] 1.00±0.001.00± 0.00 0.00±0.000.00± 0.00 909 BBBP Factual 0.50±0.030.50± 0.03 [0.49,0.50][0.49,0.50] 23.35±8.7323.35± 8.73 21.63±8.9421.63± 8.94 609 BBBP Counterfactual 0.50±0.030.50± 0.03 [0.50,0.51][0.50,0.51] 3.31±1.513.31± 1.51 1.24±1.111.24± 1.11 609 BBBP Exemplar 0.52±0.030.52± 0.03 [0.51,0.53][0.51,0.53] 5.57±1.045.57± 1.04 2.07±0.862.07± 0.86 609 TABLE XIV: Structural and probability diagnostics for the proposed method. Continuous metrics are conditional on hard success; stability is computed across explainer seeds within the same graph–model unit. Dataset View Present Removed Removed frac. |Δp|| p| Δ betw. Δ harmonic Stability MUTAG Factual 1.98±2.711.98± 2.71 17.82±6.0517.82± 6.05 0.90±0.130.90± 0.13 0.08±0.090.08± 0.09 0.22±0.040.22± 0.04 6.27±1.256.27± 1.25 0.25±0.350.25± 0.35 MUTAG Counterfactual 17.33±5.2217.33± 5.22 2.59±1.572.59± 1.57 0.13±0.070.13± 0.07 0.44±0.190.44± 0.19 0.12±0.050.12± 0.05 1.68±0.971.68± 0.97 0.74±0.290.74± 0.29 MUTAG Exemplar 14.27±4.3314.27± 4.33 5.52±1.755.52± 1.75 0.28±0.050.28± 0.05 0.08±0.080.08± 0.08 0.16±0.040.16± 0.04 3.33±0.783.33± 0.78 0.19±0.130.19± 0.13 Mutagenicity Factual 4.33±6.974.33± 6.97 26.44±16.2526.44± 16.25 0.84±0.200.84± 0.20 0.12±0.100.12± 0.10 0.21±0.050.21± 0.05 7.92±2.507.92± 2.50 0.42±0.380.42± 0.38 Mutagenicity Counterfactual 26.13±15.0826.13± 15.08 2.85±1.912.85± 1.91 0.11±0.070.11± 0.07 0.31±0.160.31± 0.16 0.10±0.060.10± 0.06 2.44±1.282.44± 1.28 0.74±0.300.74± 0.30 Mutagenicity Exemplar 24.11±15.7924.11± 15.79 6.69±1.836.69± 1.83 0.25±0.070.25± 0.07 0.08±0.060.08± 0.06 0.16±0.060.16± 0.06 4.33±0.984.33± 0.98 0.17±0.130.17± 0.13 AIDS Factual 1.24±1.351.24± 1.35 14.96±15.0314.96± 15.03 0.90±0.080.90± 0.08 0.02±0.050.02± 0.05 0.27±0.070.27± 0.07 5.13±1.885.13± 1.88 0.06±0.190.06± 0.19 AIDS Counterfactual 14.45±4.5314.45± 4.53 3.48±2.103.48± 2.10 0.19±0.090.19± 0.09 0.42±0.160.42± 0.16 0.17±0.070.17± 0.07 2.57±1.242.57± 1.24 0.94±0.140.94± 0.14 AIDS Exemplar 12.23±13.4512.23± 13.45 3.97±1.983.97± 1.98 0.29±0.060.29± 0.06 0.01±0.030.01± 0.03 0.20±0.070.20± 0.07 2.71±0.872.71± 0.87 0.24±0.100.24± 0.10 COX2_MD Factual 30.99±75.2530.99± 75.25 304.14±96.02304.14± 96.02 0.91±0.220.91± 0.22 0.07±0.070.07± 0.07 0.01±0.030.01± 0.03 21.93±6.9821.93± 6.98 0.08±0.220.08± 0.22 COX2_MD Counterfactual 313.68±51.22313.68± 51.22 4.68±2.334.68± 2.33 0.02±0.010.02± 0.01 0.10±0.110.10± 0.11 0.00±0.000.00± 0.00 0.44±0.190.44± 0.19 0.31±0.280.31± 0.28 COX2_MD Exemplar 327.15±65.22327.15± 65.22 7.97±0.347.97± 0.34 0.02±0.000.02± 0.00 0.01±0.020.01± 0.02 0.00±0.000.00± 0.00 0.47±0.040.47± 0.04 0.02±0.010.02± 0.01 BBBP Factual 2.19±2.352.19± 2.35 25.70±9.6425.70± 9.64 0.91±0.090.91± 0.09 0.08±0.070.08± 0.07 0.23±0.030.23± 0.03 7.30±1.317.30± 1.31 0.38±0.350.38± 0.35 BBBP Counterfactual 23.63±9.0223.63± 9.02 3.65±2.243.65± 2.24 0.14±0.080.14± 0.08 0.38±0.150.38± 0.15 0.14±0.060.14± 0.06 2.61±1.222.61± 1.22 0.86±0.210.86± 0.21 BBBP Exemplar 20.54±8.8520.54± 8.85 7.35±1.197.35± 1.19 0.28±0.050.28± 0.05 0.05±0.050.05± 0.05 0.19±0.040.19± 0.04 4.06±0.564.06± 0.56 0.16±0.060.16± 0.06 TABLE XV: Cross-perspective overlap for the proposed method. Overlap is the Jaccard similarity between factual retained edges and the edges deleted by the counterfactual (CF) or exemplar view. Contrast is CF minus exemplar overlap. Values are mean ± standard deviation after averaging explainer seeds within graph–model units on complete successful triplets. Counts and intervals are in the supplementary metric records. Dataset Factual–CF overlap Factual–exemplar overlap Contrast MUTAG 0.095±0.1870.095± 0.187 0.017±0.0480.017± 0.048 0.078±0.1770.078± 0.177 Mutagenicity 0.150±0.2120.150± 0.212 0.036±0.0650.036± 0.065 0.114±0.2140.114± 0.214 AIDS 0.259±0.2110.259± 0.211 0.075±0.0830.075± 0.083 0.184±0.2040.184± 0.204 COX2_MD 0.014±0.0290.014± 0.029 0.005±0.0080.005± 0.008 0.009±0.0300.009± 0.030 BBBP 0.177±0.2100.177± 0.210 0.026±0.0480.026± 0.048 0.151±0.1990.151± 0.199 The following tables report paired proposed-minus-baseline effects. Explainer seeds and predictor seeds are averaged within graph before inference; brackets are graph-level 95% bootstrap intervals and multiplicity is controlled by Holm adjustment within each dataset, architecture, perspective, and outcome family. Coverage effects use every graph. Quality effects require both methods to return a valid explanation and can therefore have very small n when a baseline rarely succeeds; these conditional effects are not population-wide quality comparisons. Values printed as p<10−300p<10^-300 reflect numerical underflow, not an exactly zero probability. TABLE XVI: Paired factual effects against compatible baselines. Dataset Baseline Coverage difference Holm p Comprehensiveness difference Holm p MUTAG GNNExplainer 0.0000.000 [0.000, 0.000]; n=188n=188 1 0.0180.018 [0.006, 0.029]; n=188n=188 0.00729 MUTAG PGExplainer 0.0000.000 [0.000, 0.000]; n=188n=188 1 0.0080.008 [-0.008, 0.024]; n=188n=188 1 MUTAG CF2 reimplementation 0.0320.032 [0.018, 0.048]; n=188n=188 0.000139 −0.400-0.400 [-0.445, -0.357]; n=188n=188 8.57e-31 MUTAG Random 0.0000.000 [0.000, 0.000]; n=188n=188 1 0.0060.006 [-0.005, 0.018]; n=188n=188 1 Mutagenicity GNNExplainer 0.0000.000 [0.000, 0.000]; n=4337n=4337 1 0.0240.024 [0.022, 0.026]; n=4337n=4337 4.79e-69 Mutagenicity PGExplainer 0.0000.000 [0.000, 0.000]; n=4337n=4337 1 0.0020.002 [-0.001, 0.005]; n=4337n=4337 0.456 Mutagenicity CF2 reimplementation 0.0320.032 [0.028, 0.035]; n=4337n=4337 4.49e-72 −0.275-0.275 [-0.282, -0.268]; n=4314n=4314 <10−300<10^-300 Mutagenicity Random 0.0000.000 [0.000, 0.000]; n=4337n=4337 1 0.0380.038 [0.036, 0.041]; n=4337n=4337 4.31e-150 AIDS GNNExplainer 0.0000.000 [0.000, 0.000]; n=1600n=1600 1 0.0010.001 [-0.000, 0.002]; n=1600n=1600 0.064 AIDS PGExplainer 0.0000.000 [0.000, 0.000]; n=1600n=1600 1 0.0050.005 [0.003, 0.006]; n=1600n=1600 9.31e-19 AIDS CF2 reimplementation 0.3830.383 [0.362, 0.405]; n=1600n=1600 4.02e-139 −0.060-0.060 [-0.068, -0.053]; n=1164n=1164 1.55e-163 AIDS Random 0.0000.000 [0.000, 0.000]; n=1600n=1600 1 0.0020.002 [0.001, 0.003]; n=1600n=1600 8.61e-06 COX2_MD GNNExplainer 0.0000.000 [0.000, 0.000]; n=303n=303 1 −0.010-0.010 [-0.019, -0.001]; n=303n=303 0.00577 COX2_MD PGExplainer 0.0000.000 [0.000, 0.000]; n=303n=303 1 −0.024-0.024 [-0.034, -0.015]; n=303n=303 8.05e-07 COX2_MD CF2 reimplementation 0.4350.435 [0.407, 0.462]; n=303n=303 2.41e-46 −0.125-0.125 [-0.150, -0.101]; n=297n=297 3.99e-42 COX2_MD Random 0.0000.000 [0.000, 0.000]; n=303n=303 1 0.0170.017 [0.010, 0.025]; n=303n=303 0.00577 BBBP GNNExplainer 0.0000.000 [0.000, 0.000]; n=203n=203 1 0.0200.020 [0.012, 0.028]; n=203n=203 0.000235 BBBP PGExplainer 0.0000.000 [0.000, 0.000]; n=203n=203 1 0.0180.018 [0.010, 0.026]; n=203n=203 0.000104 BBBP CF2 reimplementation 0.0470.047 [0.025, 0.073]; n=203n=203 0.000659 −0.426-0.426 [-0.456, -0.397]; n=199n=199 8.77e-34 BBBP Random 0.0000.000 [0.000, 0.000]; n=203n=203 1 0.0230.023 [0.015, 0.032]; n=203n=203 2.32e-05 TABLE XVII: Paired counterfactual effects against compatible baselines. Dataset Baseline Coverage difference Holm p Removed-edge difference Holm p MUTAG CF-GNNExplainer 0.6380.638 [0.580, 0.694]; n=188n=188 2.08e-26 −0.141-0.141 [-0.392, 0.088]; n=17n=17 0.256 MUTAG CF2 reimplementation 0.6030.603 [0.546, 0.660]; n=188n=188 8.94e-26 −3.370-3.370 [-3.851, -2.876]; n=47n=47 1.39e-08 MUTAG Random 0.3630.363 [0.320, 0.406]; n=188n=188 2.79e-24 −1.299-1.299 [-1.530, -1.073]; n=121n=121 2.45e-16 Mutagenicity CF-GNNExplainer 0.6150.615 [0.603, 0.628]; n=4337n=4337 <10−300<10^-300 −0.130-0.130 [-0.191, -0.070]; n=239n=239 6.06e-06 Mutagenicity CF2 reimplementation 0.6240.624 [0.612, 0.637]; n=4337n=4337 <10−300<10^-300 −3.581-3.581 [-3.889, -3.275]; n=184n=184 6.76e-29 Mutagenicity Random 0.4120.412 [0.402, 0.423]; n=4337n=4337 <10−300<10^-300 −1.694-1.694 [-1.761, -1.626]; n=1984n=1984 8.45e-263 AIDS CF-GNNExplainer 0.0460.046 [0.038, 0.055]; n=1600n=1600 1.27e-21 0.0000.000 [0.000, 0.000]; n=4n=4 1 AIDS CF2 reimplementation 0.0450.045 [0.037, 0.054]; n=1600n=1600 1.27e-21 −1.625-1.625 [-2.250, -1.000]; n=12n=12 0.00781 AIDS Random 0.0360.036 [0.029, 0.043]; n=1600n=1600 1.57e-21 −1.380-1.380 [-1.775, -1.003]; n=51n=51 7.19e-07 COX2_MD CF-GNNExplainer 0.0550.055 [0.040, 0.070]; n=303n=303 4.76e-10 – – COX2_MD CF2 reimplementation 0.0510.051 [0.038, 0.066]; n=303n=303 6.26e-10 −2.917-2.917 [-5.000, -0.833]; n=4n=4 0.125 COX2_MD Random 0.0450.045 [0.032, 0.059]; n=303n=303 8.95e-10 −2.742-2.742 [-3.894, -1.712]; n=11n=11 0.00391 BBBP CF-GNNExplainer 0.5450.545 [0.490, 0.600]; n=203n=203 3.01e-26 0.0000.000 [0.000, 0.000]; n=1n=1 1 BBBP CF2 reimplementation 0.5410.541 [0.486, 0.597]; n=203n=203 3.33e-26 −3.500-3.500 [-4.600, -2.100]; n=5n=5 0.125 BBBP Random 0.4410.441 [0.395, 0.487]; n=203n=203 3.8e-26 −2.185-2.185 [-2.591, -1.797]; n=63n=63 1.04e-10 TABLE XVIII: Paired exemplar effects against compatible baselines. Dataset Baseline Coverage difference Holm p Edit-fraction difference Holm p MUTAG Greedy control −0.011-0.011 [-0.017, -0.005]; n=188n=188 0.000941 −0.026-0.026 [-0.030, -0.021]; n=188n=188 1.6e-19 MUTAG Random 0.0440.044 [0.030, 0.059]; n=188n=188 1.35e-07 0.0110.011 [0.005, 0.017]; n=188n=188 0.000736 Mutagenicity Greedy control −0.010-0.010 [-0.011, -0.008]; n=4337n=4337 1.15e-42 −0.011-0.011 [-0.012, -0.010]; n=4337n=4337 9.02e-236 Mutagenicity Random 0.0250.025 [0.022, 0.027]; n=4337n=4337 8.28e-74 0.0150.015 [0.014, 0.016]; n=4335n=4335 1.99e-157 AIDS Greedy control −0.000-0.000 [-0.001, 0.000]; n=1600n=1600 0.102 −0.002-0.002 [-0.002, -0.001]; n=1600n=1600 7.09e-13 AIDS Random 0.0030.003 [0.002, 0.004]; n=1600n=1600 0.000112 0.0010.001 [0.000, 0.001]; n=1600n=1600 0.0013 COX2_MD Greedy control 0.0000.000 [0.000, 0.000]; n=303n=303 1 −0.000-0.000 [-0.000, 0.000]; n=303n=303 0.0682 COX2_MD Random 0.0000.000 [0.000, 0.001]; n=303n=303 0.635 0.0000.000 [-0.000, 0.000]; n=303n=303 0.145 BBBP Greedy control −0.003-0.003 [-0.006, -0.001]; n=203n=203 0.0198 −0.008-0.008 [-0.011, -0.006]; n=203n=203 8.74e-14 BBBP Random 0.0080.008 [0.003, 0.014]; n=203n=203 0.0198 0.0100.010 [0.006, 0.014]; n=203n=203 2.05e-07 Appendix E Expanded Per-Dataset Explanations Figures 7–11 enlarge the five rows of the main qualitative overview. Each strip uses one query graph and presents the factual, counterfactual, and exemplar outputs in the same left-to-right grammar as the primary BBBP fragment example. All evaluated query nodes remain visible; grey nodes have no retained incident edge but their features are still supplied to the predictor. Red dashed edges are removed and green edges retained. COX2_MD is a dense distance graph, so retained context is faint and the comparatively few removed edges are red and dashed. TU node labels denote dataset categories, not inferred chemical elements. The BBBP strip uses RDKit coordinates for its recorded SMILES but remains an edge-mask explanation; it is distinct from the separately reported sanitized fragment search. The exact primary sanitized-fragment rendering is shown in Fig. 3. Fig. 7: Expanded MUTAG query, factual, counterfactual, and exemplar case. Fig. 8: Expanded Mutagenicity query, factual, counterfactual, and exemplar case. Fig. 9: Expanded AIDS query, factual, counterfactual, and exemplar case. Fig. 10: Expanded COX2_MD query, factual, counterfactual, and exemplar case. The light complete-graph context and red dashed edits prevent the dense input from obscuring the selected changes. Fig. 11: Expanded BBBP query, factual, counterfactual, and exemplar edge-mask case using RDKit coordinates. Appendix F Reproducibility and Failure Records Every record is keyed by dataset, fold, architecture, predictor seed, explainer seed, graph ID, method, perspective, budget, and protocol hash. Atomic shards permit interrupted or parallel execution without silently overwriting completed cells. A failed explanation remains a row with its violated constraint and runtime. The claim-to-artifact map in the result root links historical generated artifacts to their sources. The submission-specific manuclaim_to_artifact_map.csv instead lists only referenced artifacts, their source records, and SHA-256 hashes, excluding unused pilot BBBP and D4 groups. The primary BBBP selection CSV and revised full-node panel manifest retain figure provenance. A separate claim_support_audit.csv records the scientific claim checks. TABLE XIX: Failure audit for the primary GCN and proposed method. Failure rate is the percentage of all attempts; the final column counts the most common machine-readable reason. Dataset View Failure rate Most common reason Count MUTAG Factual 0.0% – – MUTAG Counterfactual 32.4% No Flip Within 8 Edits 173 MUTAG Exemplar 1.1% No Nontrivial Preserving Edit Within 3 Edits 9 Mutagenicity Factual 0.0% – – Mutagenicity Counterfactual 36.3% No Flip Within 8 Edits 10406 Mutagenicity Exemplar 1.0% No Nontrivial Preserving Edit Within 8 Edits 95 AIDS Factual 0.0% – – AIDS Counterfactual 95.2% No Flip Within 3 Edits 5811 AIDS Exemplar 0.0% No Nontrivial Preserving Edit Within 4 Edits 3 COX2_MD Factual 0.0% – – COX2_MD Counterfactual 94.5% No Flip Within 8 Edits 2578 COX2_MD Exemplar 0.0% – – BBBP Factual 0.0% – – BBBP Counterfactual 45.4% No Flip Within 8 Edits 664 BBBP Exemplar 0.3% No Nontrivial Preserving Edit Within 3 Edits 2 References [1] Z. Wu, B. Ramsundar, E. N. Feinberg, J. Gomes, C. Geniesse, A. S. Pappu, K. Leswing, and V. Pande, “MoleculeNet: A benchmark for molecular machine learning,” Chemical Science, vol. 9, no. 2, p. 513–530, 2018. [2] J. Rao, S. Zheng, Y. Lu, and Y. Yang, “Quantitative evaluation of explainable graph neural networks for molecular property prediction,” Patterns, vol. 3, no. 12, p. 100628, 2022. [3] Z. Ying, D. Bourgeois, J. You, M. Zitnik, and J. Leskovec, “GNNExplainer: Generating explanations for graph neural networks,” in Advances in Neural Information Processing Systems, vol. 32, 2019. [4] D. Luo, W. Cheng, D. Xu, W. Yu, B. Zong, H. Chen, and X. Zhang, “Parameterized explainer for graph neural network,” in Advances in Neural Information Processing Systems, vol. 33, 2020, p. 19 620–19 631. [5] A. Lucic, M. A. ter Hoeve, G. Tolomei, M. de Rijke, and F. Silvestri, “CF-GNNExplainer: Counterfactual explanations for graph neural networks,” in Proceedings of the 25th International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, vol. 151, 2022, p. 4499–4511. [6] M. Bajaj, L. Chu, Z. Y. Xue, J. Pei, L. Wang, P. C.-H. Lam, and Y. Zhang, “Robust counterfactual explanations on graph neural networks,” in Advances in Neural Information Processing Systems, vol. 34, 2021, p. 5644–5655. [7] R. R. Fernández, I. M. de Diego, J. M. Moguerza, and F. Herrera, “Explanation sets: A general framework for machine learning explainability,” Information Sciences, vol. 617, p. 464–481, 2022. [8] J. Tan, S. Geng, Z. Fu, Y. Ge, S. Xu, Y. Li, and Y. Zhang, “Learning and evaluating graph neural network explanations based on counterfactual and factual reasoning,” in Proceedings of the ACM Web Conference 2022, 2022, p. 1018–1027. [9] R. García, M. A. Prado-Romero, F. Gullo, and G. Stilo, “On the minimization of graph counterfactual explanations: Theory and a local bounded search algorithm,” Machine Learning, vol. 115, 2026, article 224. [10] J. Ma, R. Guo, S. Mishra, A. Zhang, and J. Li, “CLEAR: Generative counterfactual explanations on graphs,” in Advances in Neural Information Processing Systems, vol. 35, 2022. [11] J. Chen, S. Wu, A. Gupta, and R. Ying, “D4Explainer: In-distribution explanations of graph neural network via discrete denoising diffusion,” in Advances in Neural Information Processing Systems, vol. 36, 2023. [12] D. Numeroso and D. Bacciu, “MEG: Generating molecular counterfactual explanations for deep graph networks,” in 2021 International Joint Conference on Neural Networks, 2021, p. 1–8. [13] G. P. Wellawatte, A. Seshadri, and A. D. White, “Model agnostic generation of counterfactual explanations for molecules,” Chemical Science, vol. 13, p. 3697–3705, 2022. [14] Ł. Janisiów, M. Kochańczyk, B. M. Zieliński, and T. Danel, “Enhancing chemical explainability through counterfactual masking,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 26, 2026, p. 22 155–22 163. [15] L. Chen, C. He, J. Cheng, C. Li, and Q. Guan, “Generating in-distribution counterfactual explanation for graph neural networks,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 24, 2026, p. 20 208–20 216. [16] Y. Zhang, S. B. Yang, A. Khan, and C. G. Akcora, “ATEX-CF: Attack-informed counterfactual explanations for graph neural networks,” in International Conference on Learning Representations, 2026. [Online]. Available: https://proceedings.iclr.c/paper_files/paper/2026/file/bfa008dc8caca6ebf5a047004c1c9758-Paper-Conference.pdf [17] Z. Yu and H. Gao, “MAGE: Model-level graph neural networks explanations via motif-based graph generation,” in International Conference on Learning Representations, 2025. [Online]. Available: https://proceedings.iclr.c/paper_files/paper/2025/file/7749f9c0d5f109231be21e910a3ced2-Paper-Conference.pdf [18] J. Yin, S. Wang, Z. Luo, P. Huo, H. Yan, H. Miao, C. Li, S. Pan, and C. Zhang, “Paradigm shift of GNN explainer from label space to prototypical representation space,” in International Conference on Learning Representations, 2026. [Online]. Available: https://proceedings.iclr.c/paper_files/paper/2026/file/170a27d35cd6b191b9d477f4eca4634-Paper-Conference.pdf [19] K. Amara, Z. Ying, Z. Zhang, Z. Han, Y. Zhao, Y. Shan, U. Brandes, S. Schemm, and C. Zhang, “GraphFramEx: Towards systematic evaluation of explainability methods for graph neural networks,” in Proceedings of the First Learning on Graphs Conference, ser. Proceedings of Machine Learning Research, vol. 198, 2022, p. 44:1–44:23. [20] C. Agarwal, M. Zitnik, and H. Lakkaraju, “Probing GNN explainers: A rigorous theoretical and empirical analysis of GNN explanation methods,” in Proceedings of the 25th International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, vol. 151, 2022, p. 8969–8996. [21] J. Li, M. Pang, Y. Dong, J. Jia, and B. Wang, “Graph neural network explanations are fragile,” in Proceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 235, 2024, p. 28 551–28 567. [22] S. Azzolin, S. Teso, B. Lepri, A. Passerini, and S. Malhotra, “GNN explanations that do not explain and how to find them,” in International Conference on Learning Representations, 2026. [23] RDKit contributors, “The RDKit book: Molecular sanitization,” Official software documentation, accessed October 7, 2026. [Online]. Available: https://w.rdkit.org/docs/RDKit_Book.html [24] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in International Conference on Learning Representations, 2017. [25] A. K. Debnath, R. L. Lopez de Compadre, G. Debnath, A. J. Shusterman, and C. Hansch, “Structure–activity relationship of mutagenic aromatic and heteroaromatic nitro compounds: Correlation with molecular orbital energies and hydrophobicity,” Journal of Medicinal Chemistry, vol. 34, no. 2, p. 786–797, 1991. [26] J. Kazius, R. McGuire, and R. Bursi, “Derivation and validation of toxicophores for mutagenicity prediction,” Journal of Medicinal Chemistry, vol. 48, no. 1, p. 312–320, 2005. [27] K. Riesen and H. Bunke, “IAM graph database repository for graph based pattern recognition and machine learning,” in Joint IAPR International Workshop on Structural, Syntactic, and Statistical Pattern Recognition, 2008, p. 287–297. [28] J. J. Sutherland, L. A. O’Brien, and D. F. Weaver, “Spline-fitting with a genetic algorithm: A method for developing classification structure–activity relationships,” Journal of Chemical Information and Computer Sciences, vol. 43, no. 6, p. 1906–1915, 2003. [29] N. Kriege and P. Mutzel, “Subgraph matching kernels for attributed graphs,” in Proceedings of the 29th International Conference on Machine Learning, 2012, p. 291–298. [Online]. Available: https://icml.c/2012/papers/542.pdf [30] C. Morris, N. M. Kriege, F. Bause, K. Kersting, P. Mutzel, and M. Neumann, “TUDataset: A collection of benchmark datasets for learning with graphs,” arXiv preprint arXiv:2007.08663, 2020. [31] I. F. Martins, A. L. Teixeira, L. Pinheiro, and A. O. Falcao, “A bayesian approach to in silico blood–brain barrier penetration modeling,” Journal of Chemical Information and Modeling, vol. 52, no. 6, p. 1686–1697, 2012.

Instant research discovery