{"id":"a1e847ca-2f92-495f-8256-553af9775200","arxiv_id":"2505.22252","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A benchmark built from 50,000 ChEMBL molecules shows that current explainers for graph neural networks often fail to identify the chemical substructures that define the prediction.","lead":"B-XAIC is a new benchmark of 50,000 real drug-like molecules with seven structure-detection tasks, where the correct explanation is the chemical substructure that defines each label. It tests how well graph neural network explainers recover those substructures, and finds that current methods often fail even when the models are highly accurate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.3's claim that explainers fail assumes, without evidence, that GIN's high F1 implies use of the exact target substructure; the Limitations section explicitly concedes this model confound, so the benchmark result may measure label disagreement rather than explanation unfaithfulness.","rationale":"The reader's weakest_assumption is the same as mine: high F1 does not by itself establish that the model uses the intended substructure, and the authors themselves flag this in Limitations. I agree with the CONDITIONAL verdict: the benchmark contribution is plausible and the data are real, but the strongest claim in Section 4.3 is not yet supported. The concrete intervention test I propose would settle the ambiguity: if deleting or mutating the target substructure does not change predictions, the explainers may be faithfully explaining a shortcut model, and the benchmark's central negative result about XAI would need to be reframed. I therefore recommend no change to the reader's verdict; the requested revisions should include this causal control and a random/positive-control baseline for the metrics.","tokens_in":21185,"tokens_out":5288,"duration_ms":59068,"concrete_test":"Run a causal deletion/intervention test on correctly predicted positive test molecules for each task: delete or mutate exactly the annotated target substructure (e.g., replace the B atom with carbon in the B task; replace PAINS-matched atoms with chemically similar non-matching atoms) while keeping the rest of the graph fixed, and measure the fraction of predictions that flip and the drop in predicted probability. If predictions barely change, the model solves the task via shortcuts and the explanation AUROC cannot be read as explainer failure; if predictions flip at high rates, the Section 4.3 claim survives for those tasks. Additionally, run the same explanation pipeline on a GIN trained with shuffled labels; if SE AUROC remains well above chance, the metric can spuriously align with ground-truth substructures and needs a random baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise for the Section 4.3 conclusion is that a GIN with F1>98% on each B-XAIC task makes its predictions by detecting exactly the substructure used to define the label. The paper does not test this premise; it only infers it from accuracy. The Limitations section explicitly concedes: 'We cannot definitively guarantee that the trained models base their predictions on the intended underlying chemical principles,' and states that poor explanation scores could come from the model rather than the explainer. For PAINS, rings-count, and rings-max in particular, high F1 is achievable through correlated shortcuts such as atom-type frequencies, molecular size, or generic ring counts, none of which require localizing the annotated pattern. If that is what the GIN learned, then a faithful explainer would highlight the shortcut, not the ground-truth substructure, and the low SE AUROC would measure disagreement between the explainer and the label-derived ground truth, not unfaithfulness of the explainer to the model. That would invalidate the paper's headline claim that current XAI techniques 'consistently fail to properly identify and highlight the relevant molecular structures.' The benchmark itself may still be useful, but this exact confound is what the conclusion depends on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces B-XAIC, a benchmark of 50K molecules sampled from ChEMBL with seven structure-detection tasks (single-atom presence, halogen presence, indole, PAINS, ring counting, and large-ring detection), each providing ground-truth node and edge rationales. The evaluation protocol distinguishes null explanations (NE), scored by a binary IQR outlier test, from subgraph explanations (SE), scored by AUROC. The authors evaluate eight explainers across GCN, GAT, GIN, and ProtGNN backbones and report that GIN achieves very high F1 scores on all tasks while most explainers obtain moderate or low explanation scores, especially on ring-related tasks. The paper concludes that current XAI techniques consistently fail to identify the relevant molecular structures and that B-XAIC is a needed benchmark for developing better GNN explainers.","tokens_in":21405,"tokens_out":5900,"duration_ms":65134,"significance":"If the central finding were fully established, B-XAIC would be a valuable real-world benchmark for GNN explanation evaluation, and the reported failure of existing explainers on simple substructure-recovery tasks would be an important result for the field. The dataset and code are released openly, the tasks are grounded in chemically meaningful patterns, and the NE/SE distinction is a useful design idea that goes beyond top-k thresholding. The paper also provides a broad evaluation across several architectures and explainers. However, the headline conclusion depends on an untested premise: that high predictive accuracy implies the GNN uses exactly the annotated substructure. The paper's own Limitations section concedes this confound, and the evaluation lacks random or uniform baselines and a robustness analysis of the binary NE metric. The benchmark itself remains useful as a resource, but the paper's central claim needs additional support or a more conditional formulation.","major_comments":[{"comment":"The conclusion that 'the explanations generated by the explainers consistently fail to properly identify and highlight the relevant molecular structures' rests on the assumption that a GIN with F1 above 98% on each task makes predictions by detecting exactly the substructure used to define the label. This premise is not tested. The Limitations section explicitly concedes: 'We cannot definitively guarantee that the trained models base their predictions on the intended underlying chemical principles.' For PAINS, rings-count, and rings-max in particular, high F1 is achievable through correlated shortcuts such as atom-type frequencies, molecular size, or generic ring counts, none of which require localizing the annotated pattern. If the model uses a shortcut, a faithful explainer would highlight the shortcut, not the ground-truth substructure, and the low SE AUROC would measure disagreement between the explainer and the label-derived ground truth rather than explainer unfaithfulness. I request either (i) experiments that test the premise, for example by measuring how predictions change when the target substructure is masked or removed, by training on label-permuted or shortcut-controlled splits, or by adding a random/uniform attribution baseline; or (ii) a reformulation of the conclusion as conditional on the model using the intended substructure.","section":"Section 4.3 and Section 4.2"},{"comment":"The NE metric is a binary IQR outlier test with a hand-chosen 1.5 multiplier: a prediction is scored 1 if no outliers are detected and 0 otherwise. This is a coarse, distribution-dependent measure, and no random or uniform baseline is provided for comparison. A method that outputs a perfectly flat attribution will trivially pass, while a method with slightly heavy tails but otherwise reasonable behavior may fail, and the binary pass/fail outcome discards calibration information. The choice of 1.5 is arbitrary and no sensitivity analysis is reported. I recommend reporting NE with continuous alternatives (e.g., variance, entropy, or a Kolmogorov-Smirnov test against uniformity), showing how results vary with the IQR multiplier, and including a random attribution baseline so that NE scores can be interpreted.","section":"Section 3.2 and Appendix A.4"},{"comment":"The NE/SE categorization for the counting tasks is not clearly specified and the Table 1 caption is inconsistent with the body text. The caption states 'SE denotes relevant subgraphs in positive instances, while NE denotes negative instances without explanations,' but for rings-count and rings-max the reported NE proportions are only 1.49% and 1.35%, respectively, which implies that most negative examples are assigned subgraph explanations rather than null explanations. It is not explained how the relevant substructures are defined for negative examples in these tasks, for example whether all rings are annotated in a molecule with three small rings, or how overlapping ring systems are handled. This ambiguity directly affects the construction of the SE ground truth and therefore the interpretability of the SE AUROC values for two of the seven tasks.","section":"Section 3.2 and Table 1"}],"minor_comments":[{"comment":"In the description of Alkane-Carbonyl, the text reads 'alkane (F-) and carbonyl (C=O) groups,' but the parenthetical '(F-)' appears to be a copy-paste error from Fluoride-Carbonyl; alkane groups should not be denoted with an F- label.","section":"Section 2 (Related Work)"},{"comment":"The first two tasks are described as 'Detection of organoboron and organophosphorus compounds,' but the actual task is the presence of a single boron or phosphorus atom; consider renaming these to 'boron atom presence' and 'phosphorus atom presence' for accuracy.","section":"Section 3.1"},{"comment":"The text says the tasks are 'synthetic tasks' and that GIN achieves 'near-perfect performance across our synthetic tasks'; the B-XAIC tasks use real molecular graphs with pattern-derived labels, so 'synthetic tasks' is misleading and should be replaced with something like 'structure-detection tasks.'","section":"Section 4.2"},{"comment":"The definition of NE as 'no nodes or edges are more important than others' is intuitive for single-atom tasks but needs a precise operational rule for ring-counting tasks; please specify the exact annotation procedure for negative examples in rings-count and rings-max.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful benchmark contribution, but the headline claim about explainer failure is stronger than the evidence supports, and the paper's own limitations section acknowledges the model confound. In my view the central conclusion can be repaired either by adding experiments that test whether the trained GIN actually relies on the annotated substructures, or by carefully restating the claim as a conditional benchmark result. The NE metric and the absence of random baselines also need attention before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth reading and, I think, worth publishing after some revision. The dataset is the real contribution: 50K ChEMBL molecules, seven structure-detection tasks, exact atom and edge masks per molecule, and a clean separation between null explanations (no relevant substructure) and subgraph explanations (relevant substructure). The evaluation runs nine explainers across six model variants, uses a proper significance test, and ships code and data. That is more than most XAI benchmark papers deliver, and it makes the resource immediately usable.\n\nThe headline finding—that explainers consistently fail to highlight the relevant substructures even when GIN exceeds 98% F1—is mostly supported, but not as strongly as Section 4.3 claims. The stress-test concern lands: the paper infers from high accuracy that the model \"bases its predictions on appropriate chemical principles,\" and the Limitations section concedes it cannot guarantee that. For the atomic tasks (B, P, X), the inference is nearly forced—a model cannot detect boron without using the boron atom. For rings-count and rings-max, however, shortcuts like molecular size or generic ring counts could produce high F1 without ever localizing the annotated ring. PAINS is somewhere in between, with 92.9% F1 leaving real room for the model to have learned partially different patterns. So the strongest conclusion is for the atomic tasks; the ring tasks need either a softened claim or control experiments that verify the model actually uses the target substructure. This is an addressable flaw, not a fatal one.\n\nThe NE metric is also a little rough: a binary IQR outlier test with a hand-chosen 1.5 multiplier, never validated against known uniform or non-uniform explanations, and the paper omits the trivial baselines a reviewer would want (constant uniform attribution, random attribution). Minor compared with the main confound, but worth fixing.\n\nThe circularity concern does not bother me. Ground-truth rationales derived from the same pattern definitions as the labels is a design property of any structure-detection benchmark, not a flaw.\n\nWho should read this? Anyone building or evaluating molecular GNN explainers, and anyone who needs a benchmark with exact ground-truth subgraphs. It deserves a serious referee. I would send it out, with instructions to request baselines, per-task analysis, and a revised discussion that separates what we know about atomic tasks from what remains uncertain for ring tasks.","headline":"A genuinely useful molecular explainability benchmark with a thorough evaluation, but the headline claim overreaches because the paper never shows that its high-F1 GNNs reason with the exact annotated substructures.","tokens_in":21974,"tokens_out":2718,"would_cite":true,"duration_ms":31712,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Even when a graph neural network nails a molecular task, its explanations routinely point at the wrong atoms.","keywords":["explainable AI","graph neural networks","benchmark","molecular graphs","ground-truth explanations","chemical substructures","null explanations","subgraph explanations"],"falsifier":"Train a GIN on B-XAIC tasks, take positive molecules, and mask or remove the atoms belonging to the ground-truth substructure; if the predicted probability stays high after masking, the model is using other features and the explainers' failures would not demonstrate an XAI deficiency, whereas if predictions collapse, the explainers are truly failing to find the causal substructure.","tokens_in":20957,"feed_emoji":"🧪","tokens_out":5216,"duration_ms":54326,"temperature":0.7,"pith_summary":"This paper argues that current explainers for graph neural networks are unreliable in the molecular domain: they fail to highlight the chemical substructures that actually determine a molecule's predicted label, even when the model itself is highly accurate. To make this failure measurable, the authors introduce B-XAIC, a benchmark of 50,000 small molecules from a public drug-like database, with seven structure-detection tasks whose ground-truth rationales are known atom- and edge-level masks. The evaluation separates cases where a specific subgraph is the explanation from cases where no substructure is more important than any other, allowing direct accuracy-based scoring. When nine explanation methods are tested against a GIN model that exceeds 98% F1 on every task, the best average subgraph-localization AUROC is only 0.82 at the node level and 0.65 at the edge level, with several popular methods falling near chance.","feed_headline":"Explainers miss the chemistry even for 98%-accurate GNNs","feed_subtitle":"New benchmark of 50K molecules measures whether AI explanations match the real substructures behind predictions.","key_machinery":"The benchmark's central object is the B-XAIC dataset, which pairs 50,000 real drug-like molecules with seven binary tasks defined by the presence of chemical substructures: a single boron or phosphorus atom, any halogen, an indole ring, a PAINS alert pattern, a ring count over four, and a maximum ring size over six. Each molecule carries ground-truth atom and edge explanation masks, and the evaluation splits explanations into two regimes: subgraph explanations, where a specific substructure must be highlighted and is scored by AUROC, and null explanations, where no substructure is relevant and the explanation must be uniform, scored by an interquartile-range outlier test. This two-regime design is what lets the benchmark distinguish 'wrong localization' from 'uninformative noise,' avoiding the thresholding problem of prior benchmarks.","core_discovery":"The central discovery is that explanation quality and prediction accuracy decouple in molecular GNNs: a model can be nearly perfect at a substructure-detection task while every tested explainer fails to recover that substructure. On B-XAIC, the GIN architecture reaches F1 scores above 98% across all seven tasks, yet the average subgraph-localization AUROC for the best node explainer is 0.82 and for the best edge explainer is 0.65, with many methods scoring close to chance on the null-explanation half of the benchmark. The paper attributes this systematic failure to the architectural principles of GNNs, arguing that message-passing diffusion impairs precise localization of salient features, and it positions B-XAIC as a resource to expose and address this gap.","pith_inferences":["A natural next step would be to use B-XAIC to test whether explanations improve when GNNs are trained with explicit substructure supervision or when message passing is restricted to local neighborhoods; the paper does not run such experiments.","The same evaluation design could be extended to other graph domains with known rationales, such as citation or social graphs, which would show whether the observed decoupling is chemical in origin or a general property of message-passing architectures.","Because the benchmark's tasks are all binary substructure-presence problems, a harder extension would be multi-class or multi-label tasks where overlapping rationales must be weighed, which the paper lists as ongoing work but does not validate.","The paper's own limitation suggests a concrete diagnostic: measuring how much predicted probabilities drop when the ground-truth substructure is masked would separate explainer failure from model shortcut learning, a distinction the benchmark alone cannot draw."],"forward_implications":["If correct, B-XAIC provides a ready testbed for any new XAI method targeting molecular GNNs, making it straightforward to check whether an explainer recovers known chemical substructures.","The reported decoupling implies that high task accuracy cannot be used as evidence that an explainer is faithful, so benchmarks with ground-truth rationales become necessary for validation.","The null-explanation scoring regime offers a template for evaluating explanations in situations where no single element is more important than others, which many existing evaluation metrics handle poorly.","The results suggest that post-hoc gradient and mask-based explainers, despite their popularity, may need architectural changes or additional constraints to localize substructures in message-passing networks.","The benchmark can also stress-test self-interpretable models: prototype-based GNNs underperform on ring-counting tasks, indicating that inherently interpretable designs are not automatically more faithful."],"supporting_citations":[{"why":"defines the PAINS alert substructures that constitute the ground truth for one of the seven benchmark tasks.","marker":"[5]"},{"why":"supplies the real-world drug-like molecules from which the 50,000-compound dataset is sampled.","marker":"[15]"},{"why":"GNNExplainer is one of the principal post-hoc explainers evaluated and found to fail on B-XAIC, establishing the benchmark's baseline.","marker":"[45]"},{"why":"PGExplainer, a parametric explainer, is part of the comparison set and scores near chance on edge-level null explanations.","marker":"[22]"},{"why":"GraphMask, a reinforcement-learning-based edge masker, is another evaluated explainer whose subgraph scores are low.","marker":"[31]"},{"why":"GIN is the graph architecture that reaches above 98% F1 on all tasks and is the focus of the explanation-quality analysis.","marker":"[44]"}],"fun_headline_variants":["98% accurate GNNs, but explainers miss the chemistry","GNNs ace molecular tasks, fail explanation tests","B-XAIC reveals accuracy-explanation gap in GNNs","High-scoring GNNs, low-fidelity explainers on B-XAIC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark assumes that a model with high F1 on a structure-detection task is actually predicting from the labeled substructure, so when an explainer does not highlight that substructure, the failure is attributed to the explainer rather than to the model having learned some other shortcut.","fun_headline_variants_meta":{"raw":{"variants":["98% accurate GNNs, but explainers miss the chemistry","GNNs ace molecular tasks, fail explanation tests","B-XAIC reveals accuracy-explanation gap in GNNs","High-scoring GNNs, low-fidelity explainers on B-XAIC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000767,"raw_usage":{"total_tokens":3349,"prompt_tokens":843,"completion_tokens":2506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":2430}},"tokens_in":459,"tokens_out":2506,"duration_ms":19796,"temperature":1.0,"reasoning_tokens":2430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:11:44.043640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a GIN on B-XAIC tasks, take positive molecules, and mask or remove the atoms belonging to the ground-truth substructure; if the predicted probability stays high after masking, the model is using other features and the explainers' failures would not demonstrate an XAI deficiency, whereas if predictions collapse, the explainers are truly failing to find the causal substructure.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the PAINS alert substructures that constitute the ground truth for one of the seven benchmark tasks."},{"cited_title":"Gnnexplainer: Generating explanations for graph neural networks","cited_arxiv_id":null,"evidence_quote":"GNNExplainer is one of the principal post-hoc explainers evaluated and found to fail on B-XAIC, establishing the benchmark's baseline."},{"cited_title":"Parameterized explainer for graph neural network","cited_arxiv_id":null,"evidence_quote":"PGExplainer, a parametric explainer, is part of the comparison set and scores near chance on edge-level null explanations."},{"cited_title":"Interpreting graph neural networks for NLP with differentiable edge masking","cited_arxiv_id":null,"evidence_quote":"GraphMask, a reinforcement-learning-based edge masker, is another evaluated explainer whose subgraph scores are low."}],"review_version":1}