Graph neural networks are rapidly becoming the workhorses of artificial intelligence for molecules, social networks, computer programs and infrastructure—but their reasoning can remain nearly impossible to inspect. A new study from researchers at the University of Pisa proposes a way to generate large, systematic test collections for graph explainability, potentially giving scientists a far sharper view of whether an AI explanation reflects a model’s actual decision process or merely produces a persuasive-looking guess. Published in Data Mining and Knowledge Discovery, the work introduces OpenGraphXAI, a benchmark suite containing 15 graph-classification datasets derived from real-world molecular data. The researchers also release software that can generate more than 2,000 additional benchmarks, aiming to replace a fragmented evaluation landscape with a reproducible testing framework.
The problem begins with the unusual structure of graph data. Unlike an image, which is arranged on a regular grid of pixels, or a sentence, which has a sequence of words, a graph consists of entities represented by nodes and relationships represented by edges. In a molecular graph, for example, atoms become nodes and chemical bonds become edges. A graph neural network, or GNN, learns by repeatedly passing information between neighboring nodes. At each layer, a node updates its internal representation by combining its own features with information aggregated from nearby nodes. After several rounds, the network can form a representation of the entire graph and use it to predict a property such as toxicity or anticancer activity. This flexibility has made GNNs powerful, but the same distributed calculations make it difficult to identify which parts of a graph drove a prediction.
Graph explainability methods, often grouped under the name explainable artificial intelligence, attempt to solve that problem by highlighting influential nodes, edges or subgraphs. For molecular classification, an explainer might identify a ring structure, a chemical group or a compact arrangement of atoms as the motif that supposedly caused a compound to be classified as active. Some methods learn masks that retain the most important parts of the graph; others assign relevance scores to nodes or bonds, or search for a smaller subgraph that preserves the model’s prediction. Yet judging these methods is surprisingly difficult. A visually coherent highlighted region is not necessarily the region used by the GNN. An explainer can also exploit correlations in a dataset, select redundant features or generate an explanation that looks chemically plausible while failing to represent the network’s internal logic.
The usual solution is to compare an explanation with a ground-truth motif whose importance is already known. Existing benchmarks, however, have serious constraints. Many rely on simple synthetic graphs designed around deliberately inserted patterns. Such datasets are useful for controlled experiments, but they may not reflect the complexity, noise and structural diversity of real applications. Other benchmarks contain only a small number of real-world tasks assembled by domain experts, making it difficult to draw statistically reliable conclusions. If several explainers are evaluated on just one or two datasets, a method may appear superior because it happens to suit those particular graphs. The Pisa team argues that broad collections of tasks are essential: performance should be tested across different graph sizes, motif frequencies, class imbalances and relationships between predictive patterns and background structure.
Their proposed solution uses a classic procedure from graph theory and theoretical computer science: the Weisfeiler–Leman color refinement algorithm. Despite its colorful name, the method performs a relatively simple operation. It initially assigns each node a label, or “color,” based on its attributes. During each iteration, a node updates its color according to its current label and the multiset of labels belonging to its neighbors. Nodes surrounded by different local structures gradually receive different colors. Repeating the process allows the algorithm to summarize increasingly broad neighborhoods without explicitly comparing every possible subgraph. In graph learning, the procedure is important because standard message-passing GNNs have a closely related expressive limitation: in many settings, they cannot distinguish graph structures that the Weisfeiler–Leman test also considers equivalent.
The researchers exploit that connection to mine motifs that separate graph classes while remaining learnable by GNNs. Starting with a generic graph-classification dataset, their method searches for recurring substructures associated with one class more strongly than another. Weisfeiler–Leman color refinement provides an efficient approximation for matching these subgraphs across many graphs. Candidate motifs can then serve as proxy ground-truth explanations: not absolute truths about nature, but structurally defined patterns with measurable class-discriminating power. The method is designed so that the selected motifs align with the expressive capacity of the GNN models being evaluated. This alignment matters because a benchmark would be unfair if it demanded that an explainer identify a pattern the underlying network could not, in principle, represent.
The resulting OpenGraphXAI suite is built from molecular classification datasets, including tasks connected to cancer-cell growth inhibition and toxicological assays. Some benchmark datasets originate from NCI1 and NCI109 screens, which classify small molecules according to activity against specific cancer cell lines. Others are derived from Tox21 assays involving the aryl hydrocarbon receptor pathway, the estrogen-receptor ligand-binding domain and the p53 stress-response pathway. Additional tasks come from screens involving MCF-7 breast tumor cells, MOLT-4 leukemia cells, P388 leukemia cells, PC-3 prostate cancer cells and SW-620 colon cancer cells. In each case, the original molecular graphs are transformed into graph-XAI tasks with class-associated structural motifs that can be used to test whether an explainer identifies relevant molecular regions.
The suite is intended to make evaluation more rigorous rather than to declare one universal definition of an explanation. A good explainer may need to satisfy several properties at once. It should be faithful, meaning that the highlighted structure genuinely influences the model’s output. It should be sufficiently concise to be useful to a human rather than marking most of the graph as important. It should be stable when small, irrelevant changes are made to the input, and it should generalize across examples instead of revealing only one idiosyncratic prediction. The new benchmarks allow researchers to compare such properties over many tasks. The authors report a use case in which several popular graph explainers are evaluated, illustrating how a larger benchmark collection can improve the statistical significance and interpretability of comparisons.
The technical foundation also highlights an important boundary. Weisfeiler–Leman refinement is powerful for discovering local structural distinctions, but it is not a complete test for graph equivalence and does not capture every way a graph neural network might behave. A motif identified as class-discriminating in a dataset is therefore a proxy ground truth, not proof that the motif is a causal mechanism in chemistry or biology. Real molecular activity can depend on three-dimensional conformation, stereochemistry, reaction conditions, protein binding and other information absent from a simple graph representation. The benchmarks can test whether an explainer tracks a trained model’s structural reasoning, but they cannot by themselves establish that the model has learned a scientifically correct mechanism. That distinction is crucial if these tools are eventually used in drug discovery or other high-stakes settings.
OpenGraphXAI is publicly distributed in JSON format through Kaggle, while its generation and evaluation code is hosted on GitHub. The authors’ broader goal is to make graph-XAI research easier to reproduce and harder to overinterpret. By automating benchmark construction from generic graph-classification data, the method could let researchers create controlled tests for domains beyond chemistry, including biological interaction networks, power grids, software vulnerability graphs and transportation systems. For developers of GNN explainers, the immediate benefit is a much larger testing ground; for scientists and regulators, the longer-term promise is a clearer way to ask whether an AI system’s explanation deserves trust. As graph-based AI moves deeper into medicine, materials science and security, the ability to interrogate not only what a model predicts but why it predicts it may become as important as predictive accuracy itself.
