acceptodds
Under review as a conference paper at ICLR 2027

Quantum-Inspired Dual-Path Graph Learning for Multimodal Biomedical Triadic Prediction

Abstract

Biomedical triadic prediction identifies context-dependent associations among molecular entities (e.g., RNAs, genes, and proteins), drugs, and diseases to support drug repurposing and candidate screening. However, molecular sequences, chemical structures, and textual descriptions differ in dimension and statistical properties, making direct fusion difficult. Moreover, missing modality features limit the information available for prediction, while noisy features can interfere with useful signals from other modalities. We propose Quantum-Inspired Dual-Path Graph Learning (QDGL) to integrate heterogeneous modalities and improve triadic prediction with missing or noisy modality features. First, Hilbert-sphere encoding maps modality features onto a shared complex unit sphere through learnable amplitudes and phases, providing a common space for fusion. Second, reliability-weighted fusion learns entity-specific weights to adjust the contributions of available modalities and constructs both vector representations and mixed density states. The vector path aggregates neighborhood features, while the density path propagates mixed states through state matching. Together, these paths combine available modality features and graph relations to mitigate the impact of missing features. Vector and measurement-based tensor interactions are combined for triadic association classification and disease ranking given a molecule–drug pair. Experiments on three biomedical benchmarks (ncRNADrug, CTD, and PrimeKG) show that QDGL achieves the highest mean Hit@1, MRR, and AUC among the compared methods, with maximum relative improvements of 12.04%, 8.05%, and 4.72%, respectively, over the strongest baseline for each dataset and metric. Further experiments demonstrate generalization to unseen entity pairs and robustness to missing features and structural feature perturbations, while ablations support the effectiveness of key designs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum-inspired learning, Multimodal graph learning, Density matrix representations, Biomedical triadic association prediction, Drug repurposing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_TmeeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QDGL, a model for molecule–drug–disease triadic association prediction with multimodal entity features. Each modality is mapped onto a complex unit sphere with learnable amplitudes and phases; modality states are mixed with learned reliability weights into both a real vector and a density matrix per entity. A heterogeneous GAT propagates the vectors, while a second path propagates density matrices through relation-specific unitary transforms, Hilbert–Schmidt-overlap attention and low-rank compression. Triads are scored by an MLP on vector features plus a Tucker interaction on Hermitian measurement readouts, trained with BCE, a ranking loss and four consistency losses across two graph views. On ncRNADrug, CTD and PrimeKG it reports the best Hit@1, MRR and AUC against eleven baselines.

Critical strengths
  1. The experimental program is unusually complete for this area. Beyond the random-split benchmark, the paper reports pair-overlap-stratified, pair-disjoint, and unseen-entity splits, full-candidate ranking, candidate-pool sensitivity, false-negative contamination, and representation controls (complex vector, real vector, vMF, cosine), all with five-fold SDs. Random triad splitting is explicitly flagged as mostly measuring completion within known pairs, and the harder settings are given rather than hidden.
  2. Gains are consistent across settings and metrics. In the pair-disjoint and unseen-entity evaluations, QDGL leads on nearly every cell, and full-candidate margins are larger than sampled ones (e.g. CTD molecule–drug-disjoint Hit@1 +36.86 pp over SAFKGC). The case studies attach PMIDs to 28 of 40 top-10 candidates, which is more verification than most papers in this space provide.
Critical weaknesses
  1. The gains are not attributable to the quantum-inspired path. The vector-only variant ("w/o dual path", Table 3) scores 79.46 / 95.56 / 93.65 Hit@1 against the full model's 80.20 / 96.98 / 94.34, and already beats every baseline in Table 2 by 8–18 pp. No baseline receives the same features, two-view training or six-term loss, so Table 2 cannot separate the architecture from the inputs and supervision. A plain HeteroGNN with the same features and BCE is the missing comparison.
  2. The headline benchmark is nearly saturated by construction. CTD and PrimeKG positives are closures of three pairwise relations, the graph is built from those same relations, and 71% of CTD test triads have all three pairs seen (Table S9). The hard settings tell a different story: PrimeKG PairSeen0 Hit@1 is 1.65 pp below SAFKGC, and ncRNADrug drug–disease-disjoint full-candidate Hit@1 is 13.21%. The ECFP "robustness" (−4.67 pp vs. −30 to −50 pp for controls, Table S8) coincides with Fig. S8 showing ECFP masking has almost no effect, so the model may simply not use that modality.
Questions for authors
  1. Please add a baseline with the same pretrained features, a standard HeteroGNN (e.g. the vector path alone) and BCE only, tuned with the same budget as QDGL. How much of the Table 2 gap survives?

  2. What tuning budget did each baseline receive? QDGL's settings (lr 1.78e-4, ω2 1.17, dropout 0.04/0.21/0.07) indicate an extensive search; the baseline text says only "guided by original implementations".

  3. For CTD and PrimeKG, how often is a test triad fully recoverable from the training graph (all three edges present)? Please report Table 2 restricted to PairSeen ≤ 1, since that is the setting the method is motivated by.

  4. Is ECFP used at all on ncRNADrug? Please report performance with ECFP removed entirely and compare with Table S8.

  5. Eq. 7 says U_r is parameterised by a single relation phase; App. S5.3 describes diagonal phase times Householder products. Which is implemented, and how many parameters per relation?

  6. ConvNTC reaches 3.55% Hit@1 on PrimeKG, below the 1/51 chance level for sampled ranking. Was it run correctly?

9 comments

W1. Sources of the performance gains of QDGL

We thank the reviewer for examining the ablation results closely. We agree that, under the standard random split, the variant without the density path, "w/o dual path" in Table 4 in Section 3.4, is close to the full model, with Hit@1 gaps of 0.74, 1.42, and 0.69 percentage points on the three datasets. We would like to clarify that the quantum-inspired design of QDGL spans two components, namely quantum-inspired adaptive multimodal state fusion and vector–density dual-path relational learning. This variant removes only the density path, and its vector representation is still built by the quantum-inspired state encoding and adaptive fusion, so its gap to the full model reflects only the contribution of the dual path.

The first component can be examined separately with the training setup held fixed. Replacing the quantum-inspired states with real-valued spherical states lowers Hit@1 on CTD and PrimeKG by 14.09 and 14.88 percentage points, and setting the phase to zero lowers Hit@1 on PrimeKG by 9.28 percentage points (Table 4 in Section 3.4). These variants use the same features, two-view training, and loss as the full model. In addition, following the reviewer's suggestion, HeteroGNN + BCE, trained on the same pretrained features with the BCE loss only, reaches 66.57%, 54.90%, and 54.53% Hit@1 on the three datasets, as reported in Q1.

The contribution of the second component becomes clearer when information is incomplete. With 80% of features and edges masked, the full model keeps Hit@1 near 78% on ncRNADrug, whereas both single-path variants fall to about 64% (Figure 4 in Section 3.6), and on PrimeKG the full model exceeds the vector-only variant by 23.86 percentage points (Appendix Figure S8). The two paths also correct each other. In PrimeKG Case 3 of Appendix Figure S6, the true disease ranks third under the vector readout and fifth under the quantum-inspired readout, but first under their combination.

The revised manuscript moves the comparison between the two paths under missing information into the main text as Figure 4 in Section 3.6. Table 4 in Section 3.4 reports the ablations, Appendix Figure S8 reports the full comparisons under missing information, and Appendix Figure S6 shows the path-correction cases.

W2. Saturation of the benchmark

We thank the reviewer for this observation. We agree that the construction of CTD and PrimeKG leaves many random-split test triads recoverable from known pairs, with 71.42% of CTD test triads in PairSeen3. Following the reviewer's suggestion, the main comparison in the revised manuscript is therefore reported on PairSeen ≤ 1 test triads in Table 3 in Section 3.3, as detailed in Q3. In these triads, at most one of the three pairs appears in training, and prediction relies more on the modality features of the entities, which is the situation QDGL is designed for. In this setting, the relative Hit@1 gain of QDGL over the strongest baseline rises from 6.95% on the full CTD test set to 17.58% on PairSeen ≤ 1, and from 8.45% to 19.36% on PrimeKG.

Table 3 in Section 3.3 reports the comparison on PairSeen ≤ 1 test triads, and Appendix Table S5 reports the comparison on the full test sets.

Q1. A same-feature HeteroGNN baseline trained with BCE only

We thank the reviewer for suggesting this controlled comparison. Following the suggestion, we trained HeteroGNN + BCE, a heterogeneous GNN following the HetGNN architecture [1], on the same pretrained features with the BCE loss only and tuned it on the validation split. Specifically, the search space and budget are listed below.

lrwdhetgnn_dimhiddendropoutBudgetEarly-stopping patience
[1e-5, 3e-3][1e-7, 1e-3]{64, 128, 256}{64, 96, 128}[0, 0.5]30 trials per dataset12 (ncRNADrug), 20 (CTD, PrimeKG)

The search budget of 30 trials per dataset matches that used for QDGL. The results are compared with QDGL below (mean ± SD over five folds, %).

Full test setPairSeen ≤ 1
DatasetMethodHit@1MRRAUCHit@1MRRAUC
ncRNADrugHeteroGNN + BCE66.57 ± 2.9777.24 ± 2.4568.93 ± 2.0265.04 ± 3.3676.27 ± 2.7455.93 ± 3.35
QDGL80.20 ± 0.4587.07 ± 0.3493.05 ± 0.5480.55 ± 0.6787.81 ± 0.4593.06 ± 0.81
CTDHeteroGNN + BCE54.90 ± 0.9270.04 ± 0.7793.03 ± 0.4219.64 ± 2.5334.69 ± 3.1557.74 ± 3.59
QDGL96.98 ± 0.0898.03 ± 0.0499.08 ± 0.0245.55 ± 1.8157.82 ± 1.2179.42 ± 0.51
PrimeKGHeteroGNN + BCE54.53 ± 2.7168.15 ± 1.7790.87 ± 0.4724.35 ± 3.4537.22 ± 3.4869.33 ± 2.20
QDGL94.34 ± 0.3596.18 ± 0.3498.07 ± 0.2071.26 ± 2.7379.39 ± 2.1291.20 ± 1.64

With the same pretrained features and BCE loss only, HeteroGNN + BCE still trails QDGL on all metrics. On the full test sets, HeteroGNN + BCE reaches 66.57%, 54.90%, and 54.53% Hit@1, compared with 80.20%, 96.98%, and 94.34% for QDGL. On PairSeen ≤ 1 test triads, QDGL exceeds HeteroGNN + BCE by 15.51, 25.91, and 46.91 percentage points in Hit@1. HeteroGNN + BCE stays below the strongest baselines in Appendix Table S5 on all three datasets, which reach 71.58%, 90.68%, and 86.99%, so the gap in Table 2 remains in full, with QDGL still ahead of the strongest baselines by 8.62, 6.30, and 7.35 percentage points.

HeteroGNN + BCE is now included in Table 3 in Section 3.3 and Appendix Table S5.

References

[1] C. Zhang, D. Song, C. Huang, A. Swami, and N. V. Chawla. Heterogeneous graph neural network. In KDD, pp. 793–803, 2019.

Q2. Tuning budget of the baselines

We thank the reviewer for the attention to how the baselines were tuned. We agree that the baselines should receive a tuning budget comparable to that of QDGL, and following the reviewer's suggestion, we added a hyperparameter search for each baseline. Each baseline receives 30 random-search trials on each dataset, matching the search budget of QDGL, with its original configuration counted as one trial. The search selects a configuration on the validation split of the first fold using each method's own validation metric, and the selected configuration is then retrained on all five folds and evaluated on the test sets, without using test results during the search. The search for HeteroGNN + BCE is described in Q1. The search spaces of the baselines and QDGL are listed below.

MethodLearning rateWeight decayOther hyperparameters
TransE[1e-4, 3e-3][1e-7, 1e-4]dim ∈ {128, 256, 384}, gamma ∈ [6, 18], p_norm ∈ {1, 2}
RotatE[1e-4, 3e-3][1e-7, 1e-4]dim ∈ {128, 256, 384}, gamma ∈ [6, 18]
QIQE-KGC[1e-4, 3e-3][1e-7, 1e-4]dim ∈ {100, 200, 300}, gamma ∈ [1, 8], interaction weight ∈ [0.25, 1]
QLogicE[1e-4, 3e-3][1e-7, 1e-4]dim ∈ {100, 200, 300}, gamma ∈ [6, 18], logic / TransE weights ∈ [0.5, 2]
SAGMM[1e-4, 3e-3][1e-7, 1e-4]hidden ∈ {150, 300, 450}, layers ∈ {3, 5}, dropout ∈ [0.2, 0.5]
MTGNN[1e-4, 3e-3][1e-7, 1e-4]hidden ∈ {64, 128, 192}, heads ∈ {4, 8}, common_size ∈ {64, 128}, dropout ∈ [0.2, 0.5]
SpikingHAN[1e-4, 3e-3][1e-6, 5e-4]input / temporal / scorer dropout ∈ [0.1, 0.5] / [0.2, 0.6] / [0.1, 0.3]
ConvNTC[1e-4, 3e-3][1e-7, 1e-4]rank ∈ {32, 64, 96}, channels ∈ {16, 32, 48}, alpha ∈ [0.25, 0.75], dropout ∈ [0.1, 0.5], input_dropout ∈ [0, 0.3]
SAFKGC[1e-4, 2e-3][1e-6, 1e-3]rank ∈ {250, 500, 750}, attention dropout ∈ [0, 0.3], ranking-loss weight ∈ [0.05, 0.2]
TriMoGCL[1e-4, 3e-3][1e-7, 1e-4]hidden ∈ {128, 256}, lam1 / lam2 ∈ [0.05, 0.2], tau ∈ [100, 1000]
HySAE[1e-4, 2e-3][1e-7, 1e-4]dr ∈ [0.97, 1], input_drop ∈ [0.2, 0.8], hidden / feature drop ∈ [0.1, 0.4], lamda ∈ [0.2, 0.8]
QDGL[1.2e-4, 5e-4][5e-5, 4e-4]model / edge / message dropout ∈ [0.03, 0.15] / [0.10, 0.30] / [0.05, 0.18], perturbed-view classification weight ∈ [0.90, 1.20], Siamese consistency weight ∈ [0.08, 0.20], NCE loss weight ∈ [0.01, 0.05], prototype loss weight ∈ [0.05, 0.16], density-view loss weight ∈ [0.001, 0.012]

(continued in the next comment)

Reviewer_1jpqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

I appreciate the fairly comprehensive experiments. However, I have concerns about the motivation, novelty, and presentation of the method.

(1) The motivation mainly focuses on common challenges in multimodal learning, such as noise and missing features. These problems have been widely studied, but the paper does not clearly explain the limitations of existing solutions or what makes its motivation unique.

(2) The paper does not sufficiently discuss how QDGL differs from existing methods or what advantages it offers. The related work section in the appendix mostly lists what previous methods have done, without much discussion of their limitations.

(3) The proposed method seems to be a combination of existing techniques, which is not necessarily a problem, but stronger justification is needed. Many details are also left to the appendix, so I often had to switch between the main text and the appendix to understand the method.

(4) The method uses several quantum-inspired operations, such as complex-valued states, density operators, unitary transformations, Hilbert-Schmidt overlap, and Hermitian measurements. However, it is unclear how these operations relate to biomedical entities and their relationships, or why they are particularly useful for this task.

(5) The paper uses reliability scores, reliability weights, and modality contributions somewhat interchangeably, but they are not the same. A reliable modality may contribute little to prediction, while a noisy modality may still contain useful information. It is unclear whether the learned weights really measure reliability or just reflect modality importance.

Critical strengths

The experiments appear relatively comprehensive, covering multiple baselines, ablation studies, and robustness and generalization analyses across three biomedical benchmarks.

Critical weaknesses

(1) The motivation is not very clear. The paper mainly discusses common challenges such as noisy and missing modalities, but does not explain well why existing methods cannot handle them or what makes this work different.

(2) The novelty and advantages of the proposed method are not sufficiently discussed. It remains unclear how QDGL differs from existing methods and why its quantum-inspired design is better suited to this task.

(4) The method seems to combine several existing techniques without enough justification. Also, many important details are placed in the appendix, making the main text difficult to follow.

Questions for authors

Please see the weaknesses and Paper Summary.

6 comments

Q1. Limitations of existing methods and what distinguishes our work

We thank the reviewer for pointing out that our original motivation was not clear enough. In the revised manuscript, we have clarified the motivation by explaining where missing and noisy modalities arise in this task, why existing solutions remain limited, and how QDGL addresses these limitations. Specifically, biomedical triadic prediction ranks molecule–drug–disease associations, and each entity is described by modality features drawn from separate databases. These databases cover only part of the entities and often identify the same entity by different names or identifiers [1], leaving some modalities missing and attaching others to the wrong entity. Missing modalities are known to weaken multimodal prediction [2].

Most existing methods handle such modalities during fusion by imputing missing features or by reweighting modalities with attention or modality experts [3], and then merge them into a single fused vector. In this vector, an erroneous modality becomes mixed with the correct ones in the same dimensions. For example, a drug linked by name to a similar compound receives a wrong fingerprint while its description remains correct, and its fused vector moves toward drugs with similar fingerprints. Reweighting can reduce this shift, yet as long as the fingerprint retains some weight, graph models propagate the mixed vector to linked molecules and diseases, where it can affect triadic prediction layer by layer.

QDGL addresses both limitations by keeping each modality as a separate weighted component of a mixed density state, with missing modalities masked out of the mixture. Its density path compares entities over these components, and a vector path provides complementary scores. With fixed fusion weights, an erroneous modality changes only its own terms, and the correct modalities continue to characterize the entity. Q2 gives further details of the design and its advantages.

We tested this design by shuffling drug ECFP on ncRNADrug. The Hit@1 of MM-RotatE and MM-QLogicE, which fuse adaptively weighted modalities into one vector, decreases by 34.29 and 42.79 percentage points, compared with 4.67 percentage points for QDGL (Table 5 in Section 3.5).

The motivation is revised in the Abstract and the first three paragraphs of the Introduction, and the limitations of fusion and relational learning are analyzed in Appendix S2.1 and S2.2.

References

[1] H. Zhang et al. BioMedGraphica: an all-in-one platform for joint textual biomedical prior knowledge and numeric graph generation. Bioinformatics, 42(6):btag355, 2026.

[2] M. Ma, J. Ren, L. Zhao, D. Testuggine, and X. Peng. Are multimodal transformers robust to missing modality? In CVPR, pp. 18156–18165, 2022.

[3] Y. Zhang et al. Multiple heads are better than one: Mixture of modality knowledge experts for entity representation learning. In ICLR, 2025.

Q2. Differences from existing methods and advantages of QDGL

We thank the reviewer for the concrete suggestions on the related work. We acknowledge that the original manuscript did not clearly explain where existing methods fall short in this task or how QDGL differs from them. In the revised manuscript, we have rewritten the related work to discuss the limitations of each group of methods and to compare them directly with QDGL. Specifically, QDGL differs from three groups of existing methods, each of which faces a distinct difficulty in this task.

Single-vector fusion methods merge all modalities into one fused vector, where an erroneous modality mixes with the correct ones, as discussed in Q1. QDGL instead keeps each modality as a separate weighted component of a mixed density state, and its density path compares entities through the Hilbert–Schmidt overlap, a weighted sum of similarities between individual modality states. With fixed fusion weights, a wrong fingerprint changes only its own terms, and the description of the drug still matches related entities.

Separate-channel designs can also keep modalities apart. Entity types, however, carry different modalities, and a disease that receives a fingerprint from a drug has no corresponding channel of its own, which calls for a separately designed scheme for cross-modality matching. Keeping every received modality enlarges the representation at each layer, and compressing it back with a projection or an MLP mixes the modalities again. In QDGL, aggregating neighbor states yields a mixed state of the same size, and its overlap with another state remains a weighted sum of component similarities without cross terms.

Quantum-inspired embedding methods such as ComplEx, QIQE-KGC, and QLogicE attach a state to each entity identifier and learn it only from the facts in which the entity appears, which leaves rarely observed entities with poorly trained states. QDGL builds its states from modality features, and relation-specific unitary transformations give relation-specific views of the same drug state without changing the weight of each modality component.

We confirmed these differences experimentally. The real-valued vMF density variant trails QDGL by less than two percentage points in Hit@1 on ncRNADrug with intact features, yet loses 52.66 percentage points when ECFP is shuffled, compared with 4.67 for QDGL (Table 5 in Section 3.5). Replacing the mixed state with a single pure state lowers Hit@1 on PrimeKG by 4.98 percentage points (Table 4 in Section 3.4). On test triads with at most one entity pair seen in training, QIQE-KGC and QLogicE remain below 50% Hit@1, and QDGL achieves the best results on all nine metrics (Table 3 in Section 3.3).

The revised Appendix S2 discusses the limitations of existing methods in multimodal fusion (S2.1), triadic relational learning (S2.2), and quantum-inspired representation learning (S2.3), and adds Appendix Figure S1, which compares single-vector fusion, separate-channel designs, and QDGL on a drug with a mismatched fingerprint. Section 3.5 now compares the methods under ECFP perturbation according to their fusion designs.

Q3. Relation to existing techniques, our contributions, and organization of the paper

We thank the reviewer for reading back and forth between the main text and the appendix, and we acknowledge that the organization of the original manuscript made the method harder to follow. In the revised manuscript, we have clarified which parts of QDGL build on existing techniques and which parts are new, explained why these parts are combined, and moved the key details and experiments into the main text.

We agree that part of QDGL is incremental. Complex-valued states have been used in knowledge graph embeddings and complex-valued networks, and the amplitude–phase Hilbert-sphere encoding of QDGL, which maps each modality to a unit-length complex state, is an incremental extension of these representations. The contributions of QDGL lie in how these states are fused and how they are used for relational learning.

Building on this encoding, our first contribution is quantum-inspired adaptive multimodal state fusion. QDGL combines the modality states an entity has with learned weights into one mixed density state, in which the modalities coexist as separate components whose shares equal their weights. Existing complex embeddings assign one state to each entity identifier and do not fuse multiple modalities in this way. This combination addresses the first limitation in Q1. A fused vector mixes all modalities in the same dimensions, whereas the mixed state keeps them distinguishable after fusion.

Our second contribution is vector–density dual-path relational learning, which preserves this separation during relational learning and addresses the second limitation in Q1. Relation-specific unitary transformations rotate the states for each relation without changing the weight of each modality component. The Hilbert–Schmidt overlap compares neighbors as a weighted sum of similarities between individual modality states, and with fixed fusion weights a noisy modality changes only its own terms. Hermitian measurements then read out features for measurement-based triadic scoring. A vector path based on relation-aware graph attention runs alongside, and joint scoring allows each path to compensate for errors of the other. Each technique in QDGL is therefore selected to keep the modalities distinguishable from fusion through scoring, and the two contributions together form one design.

We confirmed the contribution of each part experimentally. Removing or replacing any component lowers Hit@1 on all three datasets (Table 4 in Section 3.4). Replacing the quantum-inspired states with real-valued spherical states lowers Hit@1 on CTD and PrimeKG by 14.09 and 14.88 percentage points (Table 4), and under ECFP shuffling the real-valued vMF density variant loses 52.66 percentage points, compared with 4.67 for QDGL (Table 5 in Section 3.5). With 80% of features and edges masked on ncRNADrug, QDGL keeps Hit@1 near 78%, whereas both single-path variants fall to about 64% (Figure 4 in Section 3.6).

The revised method section gives the full construction of the relation-specific unitary transformation after its propagation equation and adds Table 1 in Section 2.2, which summarizes the role and biomedical meaning of each quantum-inspired operation. The main text now reports the state-representation and triadic-scoring controls (Figure 2 in Section 3.4), the ECFP perturbation comparison (Table 5), and the dual-path analysis under missing information (Figure 4), each placed within its own section, while the appendix retains theoretical proofs, reproducibility details, and supplementary experiments.

Reviewer_j4cPOpenReview ↗
Rating
4(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors developed QDGL, a model for predicting triad associations among a molecule, a drug, and a disease. They note that the modalities involved molecular sequences, chemical structures, and textual descriptions differ in dimension and statistics, and that entities are often missing some of them. They represent each modality on a shared complex unit sphere, and the modalities are combined per entity in two forms (a vector and a mixed-density state). These propagate along two graph paths, and their scores are combined to classify a triad and to rank candidate diseases for a molecule-drug pair. QDGL is evaluated on ncRNADrug, CTD, and PrimeKG against 11 baselines and improves on all three.

Critical strengths

The paper sets out the problem clearly, and the task of predicting context-dependent associations among molecules, drugs and diseases is well motivated for drug repurposing. The evaluation is also broad, covering three benchmarks against eleven baselines, with additional analyses of unseen entity pairs, sparse relations and missing modality features. However, I do not have sufficient background to assess this work, and I recommend assigning it to a reviewer with greater knowledge in this field.

Critical weaknesses

My mathematical background is insufficient to make a reliable judgment of the methodologies in this paper. I do have one thought about the evaluation, which I wrote as a question.

Questions for authors

In Table 3, there are many cases in which the best (full model) and second-best variant are very close, and in some cases overlap (considering the reported standard deviation), and in eight of the nine dataset–metric cells, that variant is the one with the density path removed. I wonder what the benefits would be (for example, in terms of memory and computational time) of using the second-best model instead of the best one.

1 comment

Q1. Application scenarios and benefits of QDGL and the second-best model

We thank the reviewer for this practical question. We agree that the variant without the density path ranks second in eight of the nine dataset–metric cells of Table 4 in Section 3.4 and is close to the full model in several of them. To answer the question, we compared the training and inference costs of the two models on CTD, the largest dataset. Specifically, removing the density path eliminates the propagation of mixed states and the measurement-based scoring, and the resulting costs are listed below. Training GPU hours sum training time across five folds, excluding setup and evaluation.

ModelPeak GPU memory (GB)Parameters (M)Forward GFLOPsTraining GPU hoursInference latency (ms)Inference peak GPU memory (GB)
QDGL31.23.840.320.92.660.43
w/o density path7.83.722.75.70.930.20

The two models have almost the same number of parameters, and the savings come mainly from computation and memory. The variant can therefore be useful when computing resources are the main constraint and a small loss in accuracy is acceptable, for example in rapid iterations during model development or in resource-limited settings with carefully curated inputs.

In the setting that QDGL targets, however, missing and mismatched modalities are common, and a mismatched modality often cannot be identified before prediction. We therefore recommend the full model as the default. Even under standard conditions it achieves the best results on all nine dataset–metric cells (Table 4), and its advantage grows as information becomes incomplete. With 80% of features and edges masked, it keeps Hit@1 near 78% on ncRNADrug, whereas both single-path variants fall to about 64% (Figure 4 in Section 3.6), and it exceeds the vector-only variant by 23.86 percentage points on PrimeKG (Appendix Figure S8). The two paths also correct each other. In PrimeKG Case 3 of Appendix Figure S6, the true disease ranks third under the vector readout and fifth under the quantum-inspired readout, but first under their combination. The additional cost remains moderate in absolute terms, with an inference latency of 2.66 ms and a training cost that is incurred once.

Appendix Table S3 reports the training resources of QDGL on all three datasets.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the main revisions

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their constructive comments, and we are grateful to Reviewer j4cP for engaging in the discussion and for the follow-up feedback. As the discussion period is about to close, we summarize the four main revisions below for the convenience of the Area Chair and all reviewers. The detailed responses are given under each review.

1. Clearer motivation and positioning (Reviewer 1jpq). We revised the Abstract and the first three paragraphs of the Introduction to clarify why existing methods remain limited when modalities are missing or noisy. Most methods merge all modalities into a single fused vector, where an erroneous modality mixes with the correct ones. QDGL keeps each modality as a separate weighted component of a mixed density state, so that with fixed fusion weights an erroneous modality changes only its own terms. When drug ECFP is shuffled on ncRNADrug, the Hit@1 of QDGL decreases by 4.67 percentage points, compared with 34.29 and 42.79 for MM-RotatE and MM-QLogicE (Table 5). Appendix S2 now discusses the limitations of each group of related methods and compares them directly with QDGL, with a new illustration in Appendix Figure S1. We also state clearly which part of QDGL builds on existing techniques and which parts form our two contributions, namely quantum-inspired adaptive multimodal state fusion and vector–density dual-path relational learning.

2. Biomedical meaning of the quantum-inspired operations and a self-contained main text (Reviewer 1jpq). The new Table 1 in Section 2.2 relates each quantum-inspired operation to its role in QDGL and its biomedical meaning. We unified the terminology as "modality fusion weight" and defined it as the relative contribution of each modality. The key method details and experiments, including the ECFP perturbation comparison and the dual-path analysis under missing information (Figure 4), are now in the main text.

3. Stricter evaluation and stronger baselines (Reviewer Tmee). The main comparison now also reports test triads with at most one entity pair seen in training (PairSeen ≤ 1, Table 3), where QDGL achieves the best Hit@1, MRR, and AUC on all three datasets. We added a same-feature HeteroGNN baseline trained with BCE only, and we optimized the adaptation of ConvNTC to triads of three entity types, which raises its Hit@1 on PrimeKG from 3.55% to 32.44%. QDGL remains ahead of both.

4. Sources of the performance gains and a complete method description (Reviewer Tmee). We clarified that the quantum-inspired design of QDGL spans both state fusion and dual-path relational learning. With the training setup held fixed, replacing the quantum-inspired states with real-valued spherical states lowers Hit@1 on CTD and PrimeKG by 14.09 and 14.88 percentage points, and setting the phase to zero lowers Hit@1 on PrimeKG by 9.28 percentage points (Table 4). The text following Equation (7) in Section 2.3 now gives the full composite construction of the relation unitary, with 5D = 800 trainable parameters per relation.

In addition, Appendix Table S3 reports the training resources of QDGL, following the question of Reviewer j4cP.

Thank you again for your time and effort.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.