Research map / Graph learning / GNN architectures

B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data

NeurIPS 2026 · Graph learning › GNN architectures · approach: GNN explainability · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces B-XAIC, a benchmark of 50K ChEMBL molecules with 7 substructure-based tasks and ground-truth node/edge explanations, and evaluates GNN explainers (GCN, GAT, GIN, ProtGNN) using null-explanation IQR outlier checks and subgraph-explanation AUROC.

Problem

Existing XAI evaluation for GNNs on molecules relies on synthetic datasets or simplified real-world tasks with data-derived metrics (fidelity, sparsity) that do not link to explanation faithfulness, and thresholding or top-k selection makes ground-truth-based evaluation unreliable.

Approach

A real-world molecular benchmark where ground-truth explanations are derived from chemical substructure presence, split into null explanations (no outliers should be highlighted, scored by IQR) and subgraph explanations (scored by AUROC), enabling accuracy-style evaluation of factual explainers without thresholds.

Key results

• GIN reaches F1 above 98% on all seven B-XAIC tasks (e.g., 99.96 on B, 98.32 on indole, 92.90 on PAINS), while ProtGNN variants score lower on several tasks.
• Best node-level average explanation score for GIN is about 0.67 (Saliency, with SE 0.82), and best edge-level average for GIN is 0.54 (GNNExplainer) with the GIN-PGExplainer edge average at only 0.21.
• Gradient-based explainers tend to highlight fragments even in null-explanation graphs (low NE scores), while GNNExplainer obtains the best GIN node NE score of 0.69 but lower SE.

Surprising findings

• Explainers struggle even on simple substructure tasks (boron, phosphorus, halogen) for models with near-perfect accuracy, with no universal threshold separating important nodes across methods.
• Methodologically similar explainers can produce markedly divergent explanations for the same graph.

Limitations

• Only local explanations are evaluated; global explanation support is limited.
• Models may not rely on the intended chemical features, so XAI failure cannot be cleanly separated from model failure.
• Tasks are synthetic labels defined by substructure presence on real molecules, not real-world activity labels.

Experimental setup

Datasets: ChEMBL 35

Benchmarks: B-XAIC

Metrics: F1 IQR-based null explanation score (NE) AUROC-based subgraph explanation score (SE) average of NE and SE

Baselines: GNNExplainer PGExplainer GraphMaskExplainer Saliency Integrated Gradients Input x Gradient Deconvolution GuidedBackprop ShapleyValueSampling ProtGNN

Builds on: GNNExplainer PGExplainer ProtGNN GraphMask Integrated Gradients

Closest papers by meaning

• MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation (ICLR 2025 · similarity 0.88)
• On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.87)
• Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks (ICML 2025 · similarity 0.86)
• Expressivity and Generalization: Fragment-Biases for Molecular GNNs (ICML 2024 · similarity 0.86)
• Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel (ICLR 2024 · similarity 0.85)
• Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · similarity 0.85)