acceptodds
Withdrawn from ICLR 2027

QStruct: Quantum-Circuit Structural Priors for Data-Efficient Molecular Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable molecular prediction under scarce labels and chemical distribution shift benefits from structural biases beyond local geometry. We introduce QStruct, a hybrid framework in which a pooled geometric GNN embedding sets bounded rotation angles for a shallow eight-qubit ring circuit. Eight single-qubit expectations and seven adjacent correlations form a 15-dimensional feature vector fused with the graph representation. Exact classical state-vector evaluation supplies this inductive bias without claiming quantum advantage. Across QM9, MD17, PCQM4Mv2, and OC20, reported results support data efficiency and extrapolation while retaining competitive full-data accuracy. With 100 QM9 training molecules, MAE is , versus – for strong baselines; full-data OC20 OOD-Adsorbate/Catalyst errors decrease from to . Controlled ablations favor learned circuit features over frozen or classical alternatives, while coverage curves reveal remaining calibration gaps. Separate descriptive records illustrate uncertainty ranking, computational trade-offs, and target-dependent residuals without adding independently verified benchmark evidence. Code: https://anonymous.4open.science/r/QCMRL-7203.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Quantum GNN, Molecular Prediction, Equivariant Learning, Few shot, OOD Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_1o74OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes QStruct, a hybrid molecular representation learning framework that combines a geometric GNN with a shallow parameterized quantum circuit. The circuit extracts 15 measurement features, which are fused with graph representations for molecular property prediction. Experiments on QM9, MD17, PCQM4Mv2, and OC20 report improvements in data efficiency, prediction accuracy, and generalization under distribution shift.

Critical strengths

The paper explores an interesting direction by incorporating quantum-circuit-based representations into molecular property prediction. The experiments cover several established molecular benchmarks and report promising improvements, particularly in low-data settings.

Critical weaknesses

My main concern is the poor organization and insufficient explanation of the proposed method. The technical section occupies less than one page, while most of the paper is devoted to extensive experimental results and descriptive analyses. The central motivation and technical contributions are not adequately developed. I believe the paper would benefit from substantial restructuring, with more emphasis on explaining the method rather than presenting numerous experimental figures.

The methodological novelty is also limited. QStruct mainly combines an existing geometric GNN, a parameterized quantum circuit, and feature fusion. More importantly, the paper does not sufficiently explain why the chosen circuit topology and measurement operators provide meaningful structural priors for molecular representation learning. The connection between the proposed quantum features and molecular structure remains unclear.

Another serious concern is the reliability of some experimental evidence. Section 4.4 explicitly states that the additional diagnostic records originated as illustrative layout inputs and that the original run logs are unavailable. Nevertheless, these records are presented in Figures 11–13. Although the authors distinguish them from the main benchmark results, including such unverified analyses in the experimental section raises concerns about the reliability and reproducibility of the reported findings.

Questions for authors
  1. What is the main technical contribution of QStruct beyond combining existing GNNs and parameterized quantum circuits?

  2. Why should the selected circuit topology and observables provide meaningful molecular structural priors?

  3. What evidence establishes the advantage of the quantum module over alternative structured classical representations under comparable computational budgets?

Reviewer_mxM1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes QStruct, which adds features from a small parameterized quantum circuit to a molecular graph neural network. A pooled graph embedding sets bounded rotation angles for an eight-qubit circuit. Exact classical simulation produces eight single-qubit Z expectation values and seven adjacent ZZ correlations, which are fused with the graph representation. The authors evaluate QStruct on QM9, MD17, PCQM4Mv2, and OC20, focusing on prediction with limited training data, distribution shift, and uncertainty calibration. Comparisons include no-circuit, classical bottleneck, and frozen-circuit controls. The contribution is a circuit-induced representation prior within a classical learning pipeline.

Critical strengths
  1. The fixed-size circuit branch adds a restricted family of nonlinear features while retaining the classical graph representation. This gives the authors a concrete, tractable way to test the proposed inductive bias and may be useful for molecules of different sizes.
  2. The MLP, low-rank, frozen-circuit, and no-circuit controls help assess what the additional branch contributes. On QM9 with 100 training examples, the reported improvement from 0.070 eV for PhysNet to 0.059 eV could be meaningful. These results could support a useful empirical contribution if the experiments are reproducible and adequately controlled.
Critical weaknesses
  1. The method and structural-prior ablations are not specified in enough detail to identify the source of the reported gains. Figures 7-8 and Appendix E.1 describe geometry, correlation, and hierarchy interventions without defining how they are implemented. The paper does not say whether removing correlation means deleting entangling gates, removing pairwise observables, or changing another part of the computation. The hierarchy component is also unclear. Appendix C omits the concrete widths, ranks, and parameter counts of the classical controls. These gaps prevent a reliable interpretation of the main ablation results. The MD17 force-prediction pathway is similarly unclear: the paper does not explain whether forces are energy gradients or outputs of an equivariant atomwise head.

  2. Several experimental choices limit what the data-efficiency and cross-benchmark results establish. The QM9 few-shot experiments still use 13,000 labeled validation examples for checkpoint selection and uncertainty calibration (Appendix C, lines 881-898 and 911-917). Equal access across methods supports comparisons under this protocol, but the results do not establish performance with a similarly small total label budget. For PCQM4Mv2, Table 3 includes geometry-dependent models, although the official validation set provides no reference equilibrium geometries. The paper does not explain the source of validation-time geometries or any architecture adaptations. These omissions affect the applicability and comparability of the reported results.

Questions for authors
  1. Please define the exact G/C/H interventions and provide the backbone and control configurations, including parameter counts. Which circuit specification produced the reported results, particularly given the CZ gates in Equation (1) and the controlled-X symbols in Figures 3-4?

  2. How sensitive are the main few-shot gains to the 13,000-example labeled validation set? Please provide any available evidence from smaller validation budgets, or clarify the scope of the data-efficiency claim.

  3. What geometry inputs or adaptations did each PCQM4Mv2 model use at validation time? For MD17, how are Cartesian forces computed, and how are physical units and target standardization handled?

  4. Can the authors provide paired seed-level results for the main comparisons between the circuit and classical controls, along with the configurations needed to reproduce them?

  5. How were the records underlying Figures 11-13 generated? Please distinguish actual model outputs from constructed illustrations and state what provenance is available for each figure.

Reviewer_9F2EOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

QStruct incorporates circuit feature modules into the molecular graph predictor. The authors used classical exact simulation to compute an 8-qubit, 3-layer parameterized circuit, taking the expected values of 8 Zs and 7 adjacent ZZs to form a 15-dimensional vector, which was then fused with the graph embedding. Experiments were conducted on QM9, MD17, PCQM4Mv2, and OC20. On QM9, the 100-shot MAE was 0.059, compared to 0.070 for PhysNet; the Energy MAE for full training was 0.021, on par with PaiNN and PhysNet.

Critical strengths
  1. The computational process from input to prediction is clear. Encoding, gate sequences, readout, and fusion are all mathematically defined, allowing for verification of the construction of newly added features (Eq 1–2).
  2. Feature selection is directly related to method design. Table 4 examines alternative feature constructions by replacing circuit modules with MLPs, low-rank features, and frozen features; QM9’s 100-shot MAE also outperforms the listed baselines (Table 1).
Critical weaknesses
  1. Parameter matching rules require a corresponding model configuration table. Appendix C specifies the layers, widths, pooling methods, optimizers, and sample order, with the difference in trainable parameters kept within 1% (page 18, lines 926–934). I recommend listing the number of layers, width, pooling methods, and the ratio of trainable parameters to total parameters for QStruct, BaseGNN, MLP, and the low-rank control side-by-side, while also specifying the dimensions of the fusion and prediction heads, as well as the rank of the low-rank modules. Please also indicate which published baselines follow the standard configuration. Ablations for geometry, correlation, and hierarchy also require specific implementation details: which components were removed or replaced, and which parts remained unchanged (§4.3)?

    The anonymous repository was reported as expired during peer review.

  2. The mean and error distributions for OC20 come from different training setups. Figure 6 uses 100,000 training instances, with 20,000 instances evaluated for each partition shown and each seed (lines 903–910 on page 17); Table 3, however, uses all 460,328 OC20 IS2RE training instances and the complete validation set listed. Figure 14 already shows the mean values for full-scale training (page 19). A remaining question is: after full-scale training, do the median, 90th percentile error, and difficulty-based stratification comparisons still yield the same benefits? Please supplement these results and explain why the training set was reduced in Figure 6. The training and evaluation scales should be included in the figure legends.

  3. A direct reference to the testing protocol should appear next to Figure 12. Appendix C provides the hardware and batch size on page 17, lines 881–889, and describes the timing procedure on page 18, lines 931–934. These settings should be referenced either in the main text when introducing Figure 12 or in the figure caption, briefly outlining the common conditions used for comparing error, GPU memory, and throughput.

  4. The use of validation labels should be explained just as prominently for the few-shot results. The 50/100-shot experiments used 13,000 labeled validation molecules to select checkpoints and fit calibration scale parameters (lines 891–898 on page 17). These labels were also part of the data required to obtain the reported results. It is recommended to list the number of training and validation labels side-by-side next to the results in Table 1 and Figure 5, and briefly describe their respective purposes.

  5. Figure captions require experimental background, and table formatting must be consistent. The descriptions for Figure 1, Figure 6, and Table 4 are too brief. Figure 6 should specify the dataset size and the plotted statistics; Table 4 should indicate the QM9 100-shot training, the low-similarity subset of 2,600 molecules, the eV units, and the components of each control substitution. The 0.081 in this table represents the subset MAE, while 0.059 represents the full test set MAE (§4.3); these should be labeled separately. In Table 1, the Energy MAE values for PaiNN, PhysNet, and Ours are all shown as 0.021; these should be uniformly bolded, and the rule for this formatting should be explained (page 4). The spacing between Table 4 and “Factorial Priors” should also be adjusted according to the ICLR 2027 template (page 6).

Questions for authors

Key Questions

  1. Table 4 compares modules on a subset, while Figure 7 reports ablation results on the full test set. Could you consolidate both sets of results for comparison using the same QM9 100-shot split and the full test set, and list the configurations and parameter counts alongside the MAE values?
  2. Why was the OC20 training set reduced to 100,000 entries in Figure 6? Figure 14 already shows the mean values for the full training set. I am more interested in whether the median and 90th percentile errors, as well as the difficulty-based comparisons for ID, OOD-A, and OOD-C after training on the full set, still support the same conclusions. Please include this set of results.

Secondary Questions

  1. Would the ranking of the modules in Table 4 change when evaluated on the remaining 80% of test molecules? By using the existing 100-shot models and adding the MAE values for this portion, this would clarify whether the circuit’s advantage stems primarily from molecules with low similarity.

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

Related papers

Discussion (0)

Sign in to comment.