acceptodds
Under review as a conference paper at ICLR 2027

Variational Quantum Attention for Molecular Graph Learning

Abstract

Molecular property prediction is central to computational drug discovery, where graph neural networks learn to weight neighboring atomic environments during message passing. Yet it remains unclear how variational quantum circuits alter learned attention behavior in molecular graphs. We introduce an edge-aware variational quantum attention mechanism for molecular graph learning, in which the receiving atom, neighboring atom, and connecting bond jointly determine the quantum attention state. Across five molecular property and bioactivity prediction tasks, QGAT achieves competitive performance relative to GATv2, with a consistent improvement on BBBP across all six evaluated circuit ansatzes. We further compare how the quantum and classical attention scores weight molecular structure beyond accuracy. In the Verubecestat BACE1 inhibitor series, QGAT achieves a higher Spearman correlation than GATv2 and assigns positive attributions to several structural changes consistent with reported structure-activity relationships (SARs). This case study shows that the two attention mechanisms can exhibit different prediction and attribution behavior across structurally related BACE1 analogues, while broader validation is required to determine how consistently these differences generalize across chemical series and targets. Circuit ablations further show that performance depends on the circuit design. Together, these results show that variational quantum attention can serve as a viable alternative molecular attention parameterization while inducing circuit- and chemistry-dependent behavior distinct from a matched classical scorer.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Graph Neural Networks, Quantum Attention, Molecular Property Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WDgoOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

For molecular property prediction, the authors propose an edge-aware VQC attention mechanism that extends the existing QGAT framework. Specifically, the proposed method incorporates bond attributes, which are not explicitly modeled in the original QGAT, into the quantum attention scorer. Compared with the classical GATv2, the proposed model achieves broadly comparable performance with slightly fewer parameters and outperforms GATv2 on several tasks. In addition, it exhibits attention attribution patterns that differ from those of GATv2. In several chemical case studies, these patterns are consistent with known structure–activity relationships (SARs).

Critical strengths
  • Incorporating bond attributes into quantum attention is well motivated for molecular graphs, where edge types have chemically meaningful information.
Critical weaknesses
  • The manuscript states that it adapts the existing QGAT framework of Ning et al. (2025), but it is not sufficiently clear which components are inherited from prior QGAT and which are newly introduced. As far as I can understand from the current presentation, the main methodological modification appears to be Eq. (7), namely the addition of the edge term WeeuvW_e e_{uv}. If this is indeed the primary architectural novelty, I find the methodological advance rather limited for an ICLR paper.
  • The main motivation of the paper is that prior QGAT formulations do not explicitly incorporate molecular bond attributes into the quantum attention computation. However, the manuscript does not include an otherwise identical QGAT baseline without the edge term WeeuvW_e e_{uv}. Consequently, it is unclear whether incorporating bond attributes into the quantum attention computation provides any benefit over the corresponding node-only QGAT formulation
Questions for authors
  • Which components are inherited from Ning et al. (2025), and which are newly introduced in this work?
  • Could the authors provide an ablation comparing the proposed edge-aware QGAT against an otherwise identical QGAT without the WeeuvW_e e_{uv} term?
3 comments

We thank the reviewer for this question. We would like to clarify that our primary objective is not to propose a new quantum graph attention architecture, but rather to investigate the new application of quantum attention to molecular graph learning and drug discovery. The main contribution of our study lie in the application and systematic of quantum attention for molecular property and bioactivity prediction, including investigating attention and attribution patterns in relation to experimentally reported structure-activity relationships.

For second question. We conducted an additional ablation study by removing the edge feature term WeeuvW_e e_{uv} from Equation (7). The results show that incorporating edge features improves predictive performance on four of the five tasks, including BACE, BBBP, FreeSolv and ESOL. In particular, AUC of BBBP improves fro, 0.6882 to 0.7157, while ESOL RMSE decreases from 1.0016 to 0.9103.

However, the node only variant performs slightly better on ChEMBL-BACE, achieving a Spearman of 0.7399 with 0.7357 for edge-aware model. These results suggest that incorporating bond features into quantum attention is beneficial for several molecular prediction tasks, although the advantage is task dependent rather than universal.

Thank you for conducting the additional ablation requested in my second question.

Regarding my first concern, if the main contribution is the systematic study of quantum attention in molecular graph learning and drug discovery rather than a new architecture, I think the paper should go beyond showing a few qualitative differences from GATv2. In particular, the differences in prediction and attribution behavior should be characterized more comprehensively, together with controlled empirical studies or theoretical analysis that help establish when, how, and why quantum attention differs from its classical counterpart.

Thank you for your constructive feedback. We agree that more comprehensive characterization of when and why quantum attention differs from classical attention would be valuable.

Our primary contribution is an application oriented investigation of quantum attention in molecular graph learning and drug discovery, rather than the proposal of a fundamentally new attention architecture. To this end, we evaluate predictive performance across five molecular datasets and investigate molecular attributions using SME and experimentally reported SARs. These experiments provide an initial empirical characterization of quantum attention in this application domain, while a systematic theoretical explanation of its differences from classical attention remains beyond the scope of the present study. We appreciate the suggestion and will clarify the scope and limitations of our contribution in the revised manuscript.

Reviewer_cRKeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QGAT, an edge-aware variational quantum attention mechanism for molecular graph learning that incorporates atom and bond information into quantum attention computation. Evaluated across five molecular property and bioactivity prediction tasks, QGAT achieves competitive performance against GATv2, with consistent improvements on BBBP. Additional analyses reveal distinct attention and attribution patterns that partially align with known structure–activity relationships (SARs), demonstrating the potential of quantum attention for molecular prediction.

Critical strengths
  • Introduces edge-aware variational quantum attention that incorporates both atom and bond information while retaining classical message passing
  • Evaluates multiple molecular prediction benchmarks with classical baselines and investigates six quantum circuit ansatzes through ablation studies.
  • Goes beyond predictive accuracy by analyzing attention patterns and molecular substructure attributions against experimentally reported SARs
Critical weaknesses
  • Experiments rely on classical quantum simulation, without demonstrating feasibility, scalability, or computational advantages on actual quantum hardware.
  • The chemical interpretation relies largely on qualitative case studies, without systematic quantitative evidence that quantum attention provides more chemically meaningful explanations than classical attention.
  • QGAT provides relatively modest gains over GATv2 on most benchmarks and even underperforms on FreeSolv, leaving the practical benefits of quantum attention unclear.
Questions for authors

please refer to the weakness part

2 comments

Question 1

Although evaluated by classical statevector simulation, QGAT follows a hybrid quantum–classical framework, with the VQC replacing only on the attention part rather than entire GAT.

Regarding hardware feasibility, our attention part requires only 6-7 qubits, while feature transformation, message aggregation, and prediction remain classical. Multi-head attention logits are obtained from commuting Pauli-Z observables, allowing them to be estimated from the same measurement outcomes. Previous studies have demonstrated VQC-based learning on real quantum hardwares (Ren et al., 2022; Pan et al., 2023; Zhang et al., 2026), providing experimental precedents for QML on NISQ hardware.

Regarding scalability, QGAT encodes local edge representations rather than entire molecular graphs, so its qubit requirements fixed for a given configuration. Although larger graphs still require more circuit evaluations, recent hybrid HPC developments, including NVIDIA CUDA-Q and multi-QPU/multi-GPU integration (Slysz et al., 2025), demonstrate practical infrastructure for coordinating classical and quantum workloads.

Our goal is to examine an alternative attention parameterization, not establish computational advantage; QML value need not be restricted to speedup (Schuld & Killoran, 2022). We evaluate ideal performance and simulated depolarizing and amplitude damping noise sensitivity (Appendix B.3).

Question 2

Our chemical interpretation follows established practices in molecular graph learning, where molecular attention and attribution patterns are examined against known chemical properties and experimentally reported SARs.

This chemistry grounded approach has been widely adopted in previous studies. Xiong et al. (2020) investigated learned molecular representations through attention-based visualization; Wu et al. (2023) introduced SME, demonstrated dataset level functional group attribution analysis; and Mathew et al. (2025) employed interpretable graph attention to investigate molecular features relevant to ligand optimization.

Importantly, our analysis extends beyond individual molecular visualizations. Appendix C.2 presents functional group attribution distributions across all five datasets, while Section 5.4 and Appendix D examine experimentally characterized BACE1 inhibitor series, including Verubecestat, Umibecestat, and LY3202626. In particular, our Verubecestat analysis identifies several attribution patterns consistent with experimentally reported SARs, while also acknowledging structural effects that neither model captures correctly.

While quantitative explainability benchmarks exist (Rao et al., 2022), our objective is to characterize differences in molecular weighting and attribution behavior and examine their correspondence with chemical knowledge, rather than claim universally superior interpretability. Our analysis therefore follows established chemistry-oriented evaluation practices while providing additional evidence of how quantum and classical attention differ.

Question 3

We acknowledge that the predictive gains over GATv2 are task-dependent, including a slightly lower RMSE for GATv2 on FreeSolv. However, our objective is to investigate quantum attention as a complementary component within hybrid quantum–classical learning, an active research direction in drug discovery (Smaldone et al., 2025).

Beyond predictive accuracy, interpretable molecular predictions can provide additional information for SAR analysis and lead optimization, as demonstrated in previous studies (Wu et al., 2023; Ito et al., 2025; Komissarov et al., 2025). In the Verubecestat series, QGAT achieves a higher Spearman than GATv2 (0.8264 vs. 0.7954) and assigns a positive attribution to the 5-fluoropyridine modification, consistent with reported SAR, whereas GATv2 assigns a negative attribution (Section 5.4).

These findings suggest that quantum attention may offer complementary molecular attribution patterns, potentially helping medicinal chemists identify relevant structural features and generate hypotheses for subsequent molecular optimization and experimental validation.

References

  • Ren et al. (2022). Nature Computational Science.
  • Pan et al. (2023). Nature Communications.
  • Zhang et al. (2026). npj Quantum Information
  • Slysz et al. (2025). arXiv:2508.16297.
  • Schuld and Killoran (2022). PRX Quantum, 3, 030101.
  • Xiong et al. (2020). Journal of Medicinal Chemistry.
  • Wu et al. (2023). Nature Communications.
  • Mathew et al. (2025). ACS Omega.
  • Rao et al. (2022). Patterns.
  • Smaldone et al. (2025). Chemical Reviews
  • Ito et al. (2025). Journal of Chemical Information and Modeling
  • Komissarov et al. (2025). Journal of Chemical Information and Modeling
  • Onawole et al. (2026). Journal of Cheminformatics

Thanks for the clarification. However, my concern around the predictive performance is not fully addressed and I think authors should consider iterate on the model to achieve apparent superior results on several benchmarks, which is a fundamental criteria besides other benefits of this quantum attention. Thus, I will keep my score.

Reviewer_7NQFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a new type of graph neural network that uses variational quantum circuits to learn attention weights between neighboring atoms in a molecular graph. The method is inspired by graph attention networks. To compute attention weights, the model first embeds representations of the source and target nodes along with the features of the edge between them, then constructs a multi-qubit state that serves as input to a variational circuit, which produces attention weights scaled by a learnable scale parameter per head. The model is evaluated on two classification and three regression tasks, yielding results comparable to other graph neural networks, including an analogous graph attention network architecture. Finally, the attention weights and post-hoc attributions are used to test what the model learned compared to GAT_v2.

Critical strengths
  1. The research direction is intriguing, showing how variational quantum circuits can produce attention weights in the message passing of graph neural networks. This work shows that GAT can be adapted to quantum computers without losing accuracy across five molecular property prediction tasks. The study also explores how different ansatzes influence the model performance, showing that different circuit designs work best for different tasks.
  2. The most important contribution, in my opinion, is the analysis of what the model learned. The paper shows there are differences between the proposed QGAT and GAT_v2 despite a small gap in prediction performance and similar network architecture beyond attention weight construction. It is interesting that the magnitude of attributions differs in Table 3 for the bottom three functional groups. The analysis of the BACE inhibitors is valuable and shows which patterns important for binding are seen by the model.
  3. I also appreciate that the paper is easy to follow, with a preliminaries section introducing both GAT and VQCs.
Critical weaknesses
  1. The key weakness is that the current results do not demonstrate any advantage over the existing methods. QGAT is computed using classical simulation, and although it provides some insights into how this model can be trained on quantum computers, there is currently no cost or performance advantage to this method. A brief discussion or theoretical results should be provided to support the advantage of the proposed model when real quantum hardware is accessible or further developed in the future.
  2. The second weakness is the depth of the molecular attention analysis. The leading conclusion from this work is that QGAT learns different attention scores than GAT_v2, without deeper reflection on the potential consequences. What would make the impact of this study stronger is stronger empirical evidence that the learned patterns are significantly different across multiple independent training runs. Do the learned attention scores converge toward any modes for these two models, or are differences between two runs of the same architecture just as far from each other as differences between QGAT and GAT_v2?
Questions for authors
  1. What are the practical consequences of this architecture in the future? Are there any provable benefits of QGAT, assuming the further development of quantum hardware?
  2. Are the dissimilarities between QGAT and GAT_v2 learned attention weights significantly different from what we would expect from two independent training runs of the same architecture?
  3. Are the differences between ansatzes statistically significant? How important is it to tune the circuits for a specific task?
  4. Are the lower attribution values in Table 3 reproduced under multiple training runs and observed for all ansatzes?
  5. What are the attributions on the edges in Figure 3? Which direction of message passing do they represent? Why is the nitrogen atom in the linker highlighted in red on the right molecule, even though the attribution is negative?
4 comments

Q1

Our study investigates QGAT as an alternative attention for molecular graph learning rather than calming a provable quantum advantage. Previous studies have established theoretical foundations for the expressive and learning capabilities of quantum neural networks (Abbas et al., 2021; Schuld et al., 2021), while Schuld and Killoran (2022) discuss the research value of QML beyond computational speedup.

Our ideal statevector simulations provide a noiseless reference for QGAT performance. With further development of quantum hardware, this hybrid framework could be investigated in practical molecular prediction and drug discovery workflows.

Beyond predictive accuracy, an important potential application is chemistry-grounded molecular interpretation. Previous studies have emphasized that interpretable molecular predictions can help identify structural features relevant to SAR, support lead optimization, and generate experimentally testable hypotheses (Harren et al., 2022; Wellawatte et al., 2023; Lamens & Bajorath, 2026).

In our Verubecestat analysis, QGAT identifies several attribution patterns consistent with reported SARs (Section 5.4), suggesting that quantum attention may provide complementary molecular perspectives for future medicinal chemistry investigations.

Q2

To quantitatively characterize the learned attention distributions, we computed normalized attention entropy. QGAT exhibits an entropy of 0.90, compared with 0.70 for GATv2, indicating more distributed attention weights in QGAT and relatively concentrated attention in GATv2. These results provide quantitative evidence of different attention concentration patterns between the evaluated models.

Q3

Our experiments employ shallow hardware-efficient ansatzes (HEAs), constructed from parameterized rotations and entangling gates (Kandala et al., 2017). Prior studies have shown that circuit structures influence expressibility and entangling capability (Sim et al., 2019), motivating the evaluation of different circuit configurations.

As shown in Table 2, no single ansatz performs best across all molecular prediction tasks, indicating that task-specific circuit selection is relevant. These findings motivate task-specific tuning rather than assuming a universally optimal circuit. Automated quantum architecture search provides a potential extension of this approach (Du et al., 2022).

Q4

The attribution values in Table 3 are averaged across five separately trained models corresponding to the five scaffold folds, rather than obtained from a single training run.

Our additional analysis indicates that attribution magnitudes also vary across different circuit ansatzes. Thus, the lower positive attributions to selected carbonyl-containing groups are characteristic of the reported configuration rather than universal across all ansatzes. This further highlights the importance of circuit selection, as different ansatzes can affect both predictive performance (Table 2) and molecular attribution behavior. Future work could explore automated circuit optimization through quantum architecture search (Du et al., 2022) or generative approaches such as GQE (Nakaji et al., 2024).

Q5

Thank you for pointing out the ambiguity in Figure 3. The apparent inconsistency around the linker nitrogen results from the placement of the numerical labels during visualization. The nitrogen atom has a positive IG attribution of +0.27, consistent with its red highlighting, whereas the negative value −0.45 belongs to the adjacent N14–C15 bond, not the nitrogen atom.

References

  • Abbas et al. (2021). Nature Computational Science, 1, 403–409
  • Schuld, Sweke, and Meyer (2021). Physical Review A, 103, 032430
  • Schuld and Killoran (2022). PRX Quantum, 3, 030101
  • Wellawatte et al. (2023). Journal of Chemical Theory and Computation, 19, 2149–2160
  • Lamens and Bajorath (2026). Chemical Science, 17, 1411–1422
  • Harren et al. (2022). Journal of Chemical Information and Modeling, 62, 447–462
  • Kandala et al. (2017). Nature, 549, 242–246.
  • Sim, Johnson, and Aspuru-Guzik (2019). Advanced Quantum Technologies, 2, 1900070.
  • Du et al. (2022). npj Quantum Information, 8, 62
  • Nakaji et al. (2024). The Generative Quantum Eigensolver (GQE) and Its Application for Ground State Search. arXiv:2401.09253.

Thank you for your response. I appreciate the suggested references on feasible strategies for automatically tuning circuit ansatzes. There are two remaining partially addressed questions:

Regarding response to Q1:

The predictive accuracy is comparable to other methods, and you claim applicability in chemistry-grounded interpretation. If QGAT is indeed more interpretable than other existing models that would change my score. However, interpretability is presented only on one Verubecestat experiment. Nothing in the design of the QGAT architecture seems intended to specifically improve interpretability, so why should it be more interpretable than other models? My intuition is that these results would vary across model runs and across different XAI methods. The claims about improved interpretability need to be supported by a more extensive, quantitative comparison against other methods, or by a discussion section explaining what specifically in the proposed architecture makes the model more interpretable.

Regarding response to Q2:

Thank you for providing measurements of attention entropy, which suggest that QGAT produces more distributed attention weights. However, my initial point was that these explanations may be unstable across multiple runs of the same method, and attributions can even change between two XAI methods applied to the same model. Measuring this stability and comparing across methods would confirm whether these results hold under multiple runs. Does that stability change depending on the chosen ansatz? You could also provide standard deviations for the measured normalized attention entropy.

Regarding response to Q4 (minor):

The response addresses my question, but I recommend revising the caption to explicitly say that the results are averaged across five trained models. I also recommend including standard deviations so the results capture cross-training variability.

Thank you for your constructive follow-up. We appreciate your suggestions regarding interpretability and reproducibility, and have conducted additional experiments to address these concerns within the available rebuttal period.

Regarding Q1

We apologize if our previous response gave the impression that QGAT is inherently more interpretable than classical GNNs. Our architecture was not specifically designed to improve interpretability, and we do not claim superior explanation fidelity or stability.

Rather, our goal is to investigate whether quantum attention produces complementary molecular attribution patterns that can be examined using established chemistry-grounded methods, including Integrated Gradients (IG) and Substructure Mask Explanation (SME). These techniques have been used in molecular modeling to investigate structure–activity relationships and support medicinal chemistry hypotheses.

Our Verubecestat analysis illustrates this potential application, while Appendix C.2 provides functional group attribution distributions across all five datasets. We agree that demonstrating greater interpretability than classical models would require systematic quantitative comparisons. We will clarify this distinction in the revised manuscript.

Regarding Q2

To investigate reproducibility, we conducted additional experiments.

Predictive stability: For BACE classification, we evaluated five random seeds (42–4642–46), each using 5K-fold training and evaluation, obtaining an overall ROC of 0.8185±0.01450.8185 \pm 0.0145.

Attribution variability: For ChEMBL-BACE, we examined the IG attribution of the same fluorine atom in a fixed molecule across five independently trained fold models. The attribution was 0.55±0.310.55 \pm 0.31, indicating a positive mean contribution but also noticeable cross-model variation.

Attention entropy: Our additional normalized attention entropy analysis yielded 0.980±0.0040.980 \pm 0.004, suggesting relatively consistent attention concentration across the evaluated models.

Together, these experiments provide additional quantitative evidence regarding predictive and attribution variability. Nevertheless, stable predictive performance or attention entropy does not necessarily imply stable atom-level explanations. We will clarify these limitations and report the additional results in the revision.

Regarding Q4: Table 3

Thank you for this helpful suggestion. We will revise the Table 3 caption to explicitly state that attribution values are averaged across five separately trained models corresponding to the five scaffold folds, and include standard deviations to characterize cross-training variability.

Thank you for responding to my follow-up comments. I think the choice of explainers is reasonable, and high attribution variability is not necessarily a problem with the explainers but rather a property of GNN training (and possible shortcut learning). It is interesting that the attention entropy is so high here. I wonder whether that can be explained by how gradients propagate through quantum circuits (if IG was used for explanation).

Because the other reviewers share my concerns about predictive performance and interpretability, I will keep my score, but I appreciate all the clarifications.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.