acceptodds
Withdrawn from ICLR 2027

Reasoning over Relations: Modeling Inter-Relational Structure for Drug-Drug Interaction Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Drug-drug interaction (DDI) prediction is essential for ensuring medication safety, and identifying the specific interaction type between a drug pair is of particular clinical value. Knowledge graphs (KGs) provide rich contextual information for this task. Existing methods typically model relations in KGs as fixed, type-specific representations, without encoding inter-relational structure. This limits the structural information available during entity representation learning. In this paper, we propose the Unconditioned Pair-anchored Relation Reasoning Network (UnPairRelNet), a multi-modal framework for DDI prediction. In particular, UnPairRelNet constructs a graph over relation types and generates relation representations that encode inter-relational structure, without dependence on any specific drug pair query. Subsequently, UnPairRelNet performs pair-conditioned bidirectional propagation anchored at the drug pair on the KG, with drug-pair attention aggregating local neighborhood information. The propagation reduces dependence on entity-specific parameters, naturally accommodating drugs unseen during training. In addition, UnPairRelNet encodes drug molecules at multiple granularities and fuses them with the KG representations for the final prediction. Extensive experiments on two benchmarks demonstrate that UnPairRelNet achieves the best performance on the majority of metrics and surpasses the strongest baseline by over 6% on macro F1 in scenarios involving novel drugs. The code is available at https://anonymous.4open.science/r/DDI1-03F6.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
drug-drug interaction prediction, knowledge graph, inter-relational structure modeling, relation-conditioned propagation

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_kY2ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a UnPairRelNet framework for predicting multi-class DDI prediction. The model first constructs a relation graph based on relation types, generating relation representations through degree-based partial-order connection patterns and query-agnostic message passing. It then performs bidirectional propagation across the knowledge graph, using drug pairs as anchors; as this approach avoids entity-specific parameters, it can be generalized to new drugs. Finally, the model integrates multi-granularity molecular representations to make predictions. Experiments demonstrate that the model is highly effective.

Critical strengths
  1. Demonstrates strong generalization capabilities in the context of new drug development, with comprehensive experimental coverage.
  2. The analysis is detailed, with good transparency and reproducibility.
Critical weaknesses
  1. This paper is more akin to a combination of existing components; the knowledge graph (KG) module adopts the two-stage structure from ULTRA/NBFNet, and the ablation studies are insufficient to justify the attribution of contributions. Furthermore, while the paper’s primary innovation lies in the configuration for the S1/S2 drug discovery scenarios, it lacks ablation studies specifically for these settings. Consequently, it is impossible to determine whether the performance gains stem from the relational modeling or from the multimodal components.
  2. The paper lacks an analysis regarding interpretability. In the context of relational reasoning, it fails to demonstrate whether the learned relational representations capture meaningful relational structures. Models such as EmerGNN and KnownDDI, for instance, all provide path-level explanations.
Questions for authors
  1. The title emphasizes "relational reasoning". However, the relational module merely generates fixed relation vectors based on co-occurrence graphs. What exactly is meant by "relational reasoning" in this context? Could you provide examples of reasoning paths or combinations of relations to demonstrate that the model is indeed performing reasoning?
  2. The macro F1 score for the full model in the ablation table is 0.878, while it is 0.883 in the main table; what is the source of this discrepancy? Could you also report the standard deviation for each of the ablation variants?
  3. EmerGNN employs a flow-based propagation mechanism between drug pairs to predict novel DDIs, utilizing the same S1/S2 experimental settings. How does the M3 model proposed in this paper differ from EmerGNN? Furthermore, why is EmerGNN neither cited nor included as a baseline in this study?
  4. The final prediction concatenates representations from multiple branches and feeds them into an MLP. The ablation in Table 3 is conducted only in the transductive setting. Could the authors provide ablations under S1/S2 to clarify how much each branch (KG propagation, molecular encoders, descriptors) contributes to the gains on unseen drugs?
Reviewer_j8zWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes UnPairRelNet for multi-type drug–drug interaction prediction. The method learns structural relationships among different relations through a relation graph and combines molecular information with a knowledge graph to make predictions. Its main innovations are relation encoding without prior knowledge of the target relation and bidirectional information propagation centered on the drug pair under prediction. Although the method has some novelty, the current experiments do not sufficiently demonstrate consistent advantages of its core design over simpler methods. Limitations in data splitting and ablation studies also make it difficult to determine the sources of performance gains. The current evidence is insufficient to fully support the authors’ main claims about model effectiveness and generalization to novel drugs.

Critical strengths

(1) This paper treats different relation types as independent learnable vectors. It constructs a graph with relation types as nodes and learns relation representations through the connections between nodes. This design can capture structural relationships that arise when different relations share entities. Compared with conventional relation embeddings, it can better capture latent dependencies among relation types. (2) In multi-class DDI prediction, the model must predict the interaction type of a drug pair. The target relation is usually unknown at inference time. This paper uses unconditioned relation encoding. The method generates representations of candidate relations. It does not use the true target relation. It does not depend on the target label. The design matches the multi-class DDI setting. The model can evaluate different candidate relations in a unified way at inference time. (3)The paper evaluates the model on Ryu’s dataset and the larger DrugBank dataset. It compares the model with nine representative methods, including molecular structure-based, knowledge graph-based, and multimodal methods. In addition to the standard transductive split, the authors design S1 and S2 settings with one and two unseen drugs, respectively, to assess generalization to novel drugs. The experiments report Accuracy, macro-F1, macro PR-AUC, and Cohen’s κ. They also provide the mean and standard deviation over five-fold cross-validation. Overall, the evaluation covers a broad range of metrics and settings.

Critical weaknesses

(1) Tables 6 and 8 show that the BKG edge file contains DDI edges with relation ID 0. The paper states that the background knowledge graph is fully retained in both the training and inference graphs, but it does not clearly state whether validation and test DDI edges are removed from the graph in each fold. The model relies on the global topology of the relation graph, node degrees, and multi-hop message passing. This issue directly affects all main experimental results. (2)On Ryu’s dataset, the proposed method achieves a macro-F1 of 0.883, well below MolecBioNet’s 0.912. On DrugBank, it achieves 0.830, also below MolecBioNet’s 0.846. The paper states that macro-F1 assigns equal weight to all classes. It also identifies macro-F1 as the most informative metric for ablation analysis. The proposed method does not achieve the best macro-F1 on either benchmark.The claim of overall superiority based on the best performance on most metrics lacks sufficient support. (3)Table 1 reports a macro-F1 of 0.883 for the full model on Ryu’s dataset, while Table 3 reports 0.878 for the same model. The paper does not explain this difference or clarify whether the two tables use the same data splits, cross-validation folds, random seeds, and training configurations. This inconsistency directly affects the comparability of the main and ablation experiments. (4)Both drugs in each S2 test triplet are absent from the training graph. The inference graph contains 60,893± 1,764 held-out DDI support edges, while the test set contains only 3,567±242 triplets. These support edges mainly connect unseen drugs to the existing graph. They provide substantial structural information about the supposedly novel drugs. The paper does not clearly explain how the DDI type labels of these support edges are handled. It also does not explain how test edges between unseen drugs are handled in the inference graph.

Questions for authors

(1)Do the support edges contain specific DDI types, or do they only indicate the existence of an interaction between two drugs? Do the proposed model and all baseline methods receive exactly the same support information? (2) How many unordered drug pairs and reversed drug pairs overlap across the training, validation, and test sets? Can the same drug pair have multiple DDI types? (3)Why does the full model achieve different macro-F1 scores on Ryu’s dataset in Table 1 (0.883) and Table 3 (0.878)? (4)In each cross-validation fold, does BKG relation ID 0 contain the DDI edges corresponding to the validation and test sets? Are these edges removed before constructing the relation graph, computing node degrees, and performing entity propagation?

Reviewer_7Am9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes UnPairRelNet, a multimodal framework for drug-drug interaction prediction.

It constructs a relation graph to capture relationships between different relation types and uses the learned representations to guide pair-conditioned bidirectional propagation on a knowledge graph.

Critical strengths

The paper explicitly models relationships between different relation types rather than relying on independent relation embeddings. The relation graph incorporates shared entities and degree-based structural patterns, providing a concrete way to capture inter-relational information.

The method considers generalization to unseen drugs by reducing dependence on entity-specific parameters during propagation and using molecular descriptors when learned drug embeddings are unavailable. This addresses an important practical challenge in DDI prediction.

Critical weaknesses
  1. Relation-level graph construction and message passing have already been explored in prior work, particularly ULTRA (ICLR 2024). The paper does not sufficiently establish what fundamentally new reasoning mechanism is introduced beyond adapting existing techniques to DDI prediction.

  2. The ablation studies demonstrate the usefulness of KG propagation, but do not isolate the contribution of the proposed relation graph construction and encoding. It makes the reviwer unclear how much of the improvement comes from these specific designs rather than existing graph reasoning mechanisms.

Questions for authors
  1. How does the proposed relation reasoning fundamentally differ from ULTRA (ICLR 2024)?

  2. Can the authors isolate the contribution of the proposed relation graph encoding? In particular, how does it compare with fixed relation embeddings, simple relation co-occurrence graphs, or an existing relation graph encoder under the same propagation framework?

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

Related papers

Discussion (0)

Sign in to comment.