Pith. sign in

REVIEW 3 major objections 4 minor 71 references

Towards Interpretable Drug-Drug Interaction Prediction: A Graph-Based Approach with Molecular and Network-Level Explanations

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MolecBioNet predicts drug-drug interaction types by treating the pair as a unified entity on a merged biomedical-molecular graph, and its two pooling mechanisms supply mechanistic explanations for each prediction.

desk verdict A solid, incrementally novel DDI model with real benchmark gains, but the cold-start protocol likely leaks test labels and the MI regularizer is a mathematical no-op; worth refereeing if the authors can fix both. read the letter →

arxiv 2507.09173 v1 pith:O5RXG3EA submitted 2025-07-12 cs.LG cs.AIq-bio.MN

Mengjie Chen , Ming Zhang , Cunquan Qu This is my paper
classification cs.LGcs.AIq-bio.MN
keywords drug-druginteractionpredictiongraphneuralnetworksbiomedicalknowledgemolecularsubstructuresinterpretablepoolingcold-startdrugmutualinformationminimizationmulti-scalerepresentationlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Drug-drug interaction prediction usually treats each drug as a separate entity, encodes it, and concatenates the two outputs; this paper argues that such representations miss the context-dependent nature of the interaction itself. MolecBioNet instead makes the drug pair the atomic modelling unit: it extracts a $k$-hop local subgraph around the pair in a merged biomedical knowledge graph and builds a hierarchical interaction graph from substructure-level molecular graphs. The paper claims that this pair-centered, multi-scale design predicts interaction types more accurately than the compared state-of-the-art methods on Ryu's and DrugBank datasets, including cold-start settings for drugs with little or no historical interaction data. It further claims that the two customized pooling mechanisms — context-aware subgraph pooling (CASPool) and attention-guided influence pooling (AGIPool) — make the predictions interpretable by naming the biological entities and molecular substructures that drive each interaction. If correct, this offers a path to DDI risk screening that both scores pairs and points at the mechanism.

What carries the argument

The load-bearing object is a pair-centered dual-graph representation. The first graph is the task-specific biomedical knowledge graph (tsBKG), built by merging a DDI graph with an external knowledge graph; for each drug pair the model takes the $k$-hop enclosing subgraph, encodes it with a Graph Transformer, and reads it out with CASPool, which computes an attention score for every entity against the concatenated embedding of the two drugs. The second graph is the hierarchical interaction graph (HIG), obtained by BRICS fragmentation of each drug's molecular graph into substructure nodes; a GCN propagates information along intra-drug bonds and a Graph Attention Network along inter-drug edges, and AGIPool aggregates substructure embeddings weighted by incoming attention. The machinery also includes a mutual-information-minimization loss on the pooled embeddings and a center loss on drug embeddings. These pooling mechanisms are simultaneously the accuracy boosters and the explanation channel, since their attention weights indicate which biological entities and which chemical substructures matter for a given prediction.

What would settle it

Re-run the Novel Drug-Existing Drug and Novel Drug-Novel Drug experiments with the task-specific biomedical knowledge graph rebuilt and all node embeddings retrained from scratch inside every fold after deleting the test-fold DDI edges; if accuracy and F1 fall to the level of the baselines, the cold-start advantage comes from the model having already seen the test interactions, and if the gap persists, the claim is confirmed.

Watch

Extended reading notes

Core claim

MolecBioNet's central discovery is that the pair itself, not the individual drug, should be the unit of representation for drug-drug interaction prediction. On the biological side, it induces a local subgraph centered on the pair $(u,v)$ inside a task-specific biomedical knowledge graph, augments each node with position and entity-type encodings, and pools the subgraph with CASPool, which attends to entities most relevant to the pair. On the molecular side, it fragments each drug with the BRICS rules into substructure nodes, connects the two drugs' substructure graphs into a hierarchical interaction graph with intra-drug and inter-drug edges, and pools with AGIPool, which scores each substructure by the attention it receives from neighbours. A mutual-information-minimization regularizer keeps these two views complementary rather than redundant, and a center loss stabilizes the per-drug embeddings across pairs. Experimental results on Ryu's dataset and DrugBank 6.0 show that the model outperforms all compared baselines across accuracy, F1, PR-AUC, and Cohen's kappa, and the cold-start experiments show the advantage persists for drugs with little or no historical interaction data. The pooling attention weights double as mechanistic explanations, identifying chemical substructures and biological entities that drive each prediction.

Load-bearing premise

The load-bearing premise is that the node embeddings used in the cold-start experiments were not already informed by the held-out interaction edges under test, since the paper does not state that the knowledge graph and GraphSAGE embeddings are recomputed per fold with test edges removed.

Editorial extensions

If this is right

  • A pair-centered representation should become the default template for DDI models, because it captures inter-drug dependencies that separate encoding and concatenation cannot express.
  • The attention weights of AGIPool can be read as ranked hypotheses about which chemical fragments mediate an interaction, giving medicinal chemists a concrete list to test before committing to experiments.
  • The CASPool weights tie predictions to named proteins, pathways, and side-effect nodes, so a predicted interaction can be interrogated for biological plausibility rather than accepted on trust.
  • Because molecular substructure information enters the embedding directly, the model should transfer to drugs absent from historical DDI databases better than network-only baselines, which the cold-start experiments indicate.
  • The mutual-information-minimization objective predicts that fusion gains come from complementary views; removing the term should hurt accuracy most when the two graphs carry overlapping signal, and the ablation results are consistent with that.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the cold-start results would be conclusive only if the biomedical knowledge graph and its GraphSAGE embeddings are rebuilt inside each fold with the test DDI edges removed; the paper does not state this, so the novel-drug advantage should be treated as provisional until that setup is described.
  • Editorial inference: the same pair-as-entity architecture could be ported to drug-food, drug-herb, or drug-metabolite interaction prediction, where the two-view (biological context plus molecular structure) design should transfer directly.
  • Editorial inference: the AGIPool influence scores could be validated quantitatively against CYP inhibition databases, effectively turning the explanation channel into a mechanistic prediction model that outputs the enzyme a fragment is likely to engage.
  • Editorial inference: the mutual-information-minimization regularizer is a general recipe for any multimodal prediction task where two views of one entity must stay complementary, not just for DDI prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes MolecBioNet, a graph-based framework for multi-class drug-drug interaction (DDI) prediction. The model builds a task-specific biomedical knowledge graph by merging the DDI graph with an external biomedical knowledge graph, extracts k-hop subgraphs around each drug pair, and encodes them with a Graph Transformer followed by a context-aware subgraph pooling module (CASPool). In parallel, it fragments each drug into BRICS substructures, builds a hierarchical interaction graph with intra- and inter-drug edges, and pools the resulting node embeddings with an attention-guided influence pooling module (AGIPool). The two representations are concatenated with per-drug biological and molecular embeddings, and the model is trained with cross-entropy, center loss, and a mutual-information-based regularization term. Experiments on Ryu's dataset and DrugBank report state-of-the-art accuracy, F1, PR-AUC, and Cohen's kappa, as well as cold-start experiments for novel drugs, ablations, and interpretability analyses including fidelity scores and literature-matched substructure hits.

Significance. If the empirical claims hold, the paper would make a useful contribution to DDI prediction by combining network-level and molecular-level context in a single interpretable framework. Strengths of the submission include the public code repository, the use of corrected paired t-tests against the second-best baseline in the main comparison, and the explicit attempt to evaluate cold-start generalization and explainability. However, the two load-bearing technical issues identified below—the likely transductive leakage in the cold-start protocol and the invalid mutual-information loss—mean that the current manuscript does not yet establish the headline claims. The main benchmark comparison may still stand after correction, but the experimental protocol and the InfoMin formulation need substantial revision.

major comments (3)
  1. [§4.2 and §3.1.1] The cold-start evaluation does not control for transductive leakage. The task-specific biomedical knowledge graph is defined as G_BKG = G_DDI ∪ G_KG (§2.2), and the GraphSAGE node embeddings x_v^(L1) are computed on the full G_BKG before subgraph extraction (§3.1.1). Section 4.2 does not state that test-fold DDI edges are removed from G_BKG, or that embeddings are recomputed per fold, before the Novel Drug–Existing Drug and Novel Drug–Novel Drug splits. Since a held-out drug's test interactions are present when x_v^(L1) is generated, the gains in Table 2 can be explained by transductive interpolation rather than by generalization to genuinely novel drugs. Please specify the exact masking protocol and, if masking was not performed, rerun the experiments with per-fold masking of all edges incident to test drugs (or at least test DDI edges) before computing embeddings.
  2. [§3.3, Eqs. (14)–(18)] The mutual-information loss is not a valid minimization of MI as written. Substituting Eqs. (16) and (17) into Eq. (15) returns exactly Eq. (14), so Eq. (15) is an algebraic identity rather than a tractable reformulation. Moreover, Eq. (18) drops the joint-entropy term H(h,z); under the same definitions the loss reduces to E[H(h)+H(z)]. This is not MI but an upper bound on it, so minimizing it can encourage entropy collapse rather than information diversity. The paper also provides no estimator for H(h) or H(z), and h_u,v and z_u,v are deterministic functions, so their differential entropy is not well-defined without an additional probabilistic model. The InfoMin term should be replaced by a proper MI estimator (e.g., InfoNCE or MINE) or explicitly presented as a heuristic decorrelation regularizer, and the wording in the abstract and contributions should be adjusted accordingly.
  3. [§4.4, Table 3] The fidelity-based interpretability evidence is not conclusive. Fidelity+ decreases from 0.282 at sparsity 0.5 to 0.213 at sparsity 0.9, meaning that removing a growing number of the 'most important' nodes produces a smaller drop in predictive fidelity; this is the opposite of what one expects from an increasingly destructive mask, and the paper's reading ('explanations gradually become less necessary') is not the standard interpretation. The evaluation lacks a control (e.g., random node masking or a baseline explainer) and reports no significance tests. Since the interpretability contribution is one of the paper's three stated pillars, please provide a control comparison and a more careful interpretation, or temper the faithfulness claim.
minor comments (4)
  1. [§4.2] The protocol for choosing 'novel' drugs is not described: please state how many drugs are selected per fold, whether they appear in the tsBKG as nodes, and how the K-fold split is constructed so that the experiments are reproducible.
  2. [Table 2] No p-values or confidence intervals are reported for the cold-start comparisons; because some absolute margins are small (e.g., Novel Drug–Existing Drug ACC 0.652 vs. 0.636), statistical significance should be assessed.
  3. [Figure 2] Figure 2 is very dense and the subfigure typography is difficult to read; a larger version with clearly labeled panels (a)–(d) would help the reader follow the method.
  4. [§4.4] The case study refers to figures in the appendix; the main text should state that Fig. 5 and Fig. 6 appear in Appendix C so that readers can locate them.

Circularity Check

1 steps flagged · score 3.0 of 10

Main benchmark comparison is independent; the MI-regularization loss is a definitional collapse, and the cold-start setup is under-specified regarding per-fold graph masking.

  1. other [Section 3.3, Eqs. (14)-(18)]
    "Since directly computing H(h_u,v) and H(z_u,v) is challenging, we reformulate MI(h_u,v, z_u,v) using conditional entropy and KL divergence: ... KL(h_u,v∥z_u,v)=H_{z_u,v}(h_u,v)-H(h_u,v), ... KL(z_u,v∥h_u,v)=H_{h_u,v}(z_u,v)-H(z_u,v). ... Since H(h_u,v, z_u,v) is non-negative, the mutual information loss is defined as: L_MI = E_{(u,v)∼D}[ H_{z_u,v}(h_u,v)+H_{h_u,v}(z_u,v)-KL(h_u,v∥z_u,v)-KL(z_u,v∥h_u,v)]."

    Using the paper's own Eqs. (16)-(17), the two KL terms in the L_MI expression are H_z(h)-H(h) and H_h(z)-H(z). Substitution cancels H_z(h) and H_h(z), leaving L_MI = E[H(h)+H(z)]: the joint-entropy term H(h,z) was dropped from Eq. (15) without a replacement, so the quantity minimized is the sum of marginal entropies, not mutual information and not a function of the h-z dependence. Eq. (15) itself is just Eq. (14) rewritten, so the 'reformulation' contributes no new computable content. Consequently the claim that this loss 'constrains the embeddings h_u,v and z_u,v to capture unique and complementary aspects' is attached, by definition, to an objective that has lost the MI term; the regularization's stated effect does not follow from the loss as written.

full rationale

The central accuracy claim in Table 1 is a direct empirical comparison against eight external baselines on Ryu's and DrugBank datasets, with standard deviations and corrected paired t-test p-values; it does not reduce the target into the input, and no load-bearing self-citation is present. The only in-text derivation that collapses is the MI regularization: Eq. (15) is algebraically identical to Eq. (14), and the implemented L_MI in Eq. (18) reduces to marginal entropies once Eqs. (16)-(17) are substituted, so the named MI objective is not MI. This is a genuine component-level flaw but does not by itself invalidate the main benchmark comparison. The cold-start experiments (Section 4.2) carry a separate rigor risk: the tsBKG is formally defined as E_BKG = E_DDI ∪ E_KG on the full dataset, and Section 3.1.2 uses 'precomputed embeddings' x_i^(L1), with no statement that G_BKG and the embeddings are rebuilt per fold after removing test DDI edges. If they are not masked, the novel-drug numbers would be transductive leakage rather than generalization; the manuscript's reference to a public repository for the exact setup leaves this unresolved. I weight this as missing support rather than confirmed circularity, keeping the score at 3.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard GNN machinery, a domain assumption about substructure-level chemistry, and the paper-specific MI loss derivation that is mathematically inconsistent. The main free parameters (k, beta, gamma, architecture sizes) are chosen by sensitivity analysis or left unspecified. No new entities are postulated.

free parameters (4)
  • beta (center loss weight) = 2
    Chosen by sensitivity analysis (Fig. 4A) that tunes beta in {0.5..3} on F1; value 2 gives peak performance.
  • gamma (InfoMin loss weight) = not explicitly stated
    Appendix B says gamma is scaled so that prediction loss is about ten times the MI loss; exact value and the MI estimator are not given.
  • k (subgraph hop count) = 2
    Chosen by sensitivity analysis (Fig. 4B) comparing k=1,2,3; 2 is best.
  • GNN layer depths L1-L4, hidden dimensions, attention heads = not reported
    Hyperparameters for GraphSAGE, Graph Transformer, GCN and GAT modules are not listed in the paper, so the architecture cannot be exactly reproduced without the code.
assumptions (5)
  • standard math Standard GNN update rules: GraphSAGE, GCN, GAT, Graph Transformer equations are taken from cited prior work.
    Sections 3.1.1-3.2.1 rely on established definitions without re-derivation.
  • domain assumption BRICS fragmentation produces chemically meaningful substructures that are the correct granularity for explaining DDIs.
    Section 2.3 and 4.4 assume functional-group-level nodes capture interaction mechanisms, following refs [7,15,61].
  • domain assumption A k-hop enclosing subgraph of the merged biomedical knowledge graph contains the biological context relevant to a DDI.
    Section 3.1.2 assumes influence is localized to k hops, citing [56].
  • ad hoc to paper GraphSAGE embeddings computed on the full tsBKG do not leak test-fold DDI information in the cold-start experiments; the paper does not state that the graph is masked per fold.
    Section 4.2 reports novel-drug results without describing split construction or embedding recomputation.
  • ad hoc to paper The mutual information identity in Eq. 15 is valid under the paper's definitions of H_z(h), H_h(z) and KL.
    Section 3.3 defines H_z(h) as conditional entropy but then uses it in KL definitions as if it were cross-entropy; the resulting loss is not MI.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Interpretable Drug-Drug Interaction Prediction: A Graph-Based Approach with Molecular and Network-Level Explanations." pith.science (2026). https://pith.science/paper/O5RXG3EA

@misc{pith2026250709173,
  author       = {Pith},
  title        = {Pith review of: Towards Interpretable Drug-Drug Interaction Prediction: A Graph-Based Approach with Molecular and Network-Level Explanations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O5RXG3EA}},
  note         = {Machine review of arXiv:2507.09173}
}
read the original abstract

Drug-drug interactions (DDIs) represent a critical challenge in pharmacology, often leading to adverse drug reactions with significant implications for patient safety and healthcare outcomes. While graph-based methods have achieved strong predictive performance, most approaches treat drug pairs independently, overlooking the complex, context-dependent interactions unique to drug pairs. Additionally, these models struggle to integrate biological interaction networks and molecular-level structures to provide meaningful mechanistic insights. In this study, we propose MolecBioNet, a novel graph-based framework that integrates molecular and biomedical knowledge for robust and interpretable DDI prediction. By modeling drug pairs as unified entities, MolecBioNet captures both macro-level biological interactions and micro-level molecular influences, offering a comprehensive perspective on DDIs. The framework extracts local subgraphs from biomedical knowledge graphs and constructs hierarchical interaction graphs from molecular representations, leveraging classical graph neural network methods to learn multi-scale representations of drug pairs. To enhance accuracy and interpretability, MolecBioNet introduces two domain-specific pooling strategies: context-aware subgraph pooling (CASPool), which emphasizes biologically relevant entities, and attention-guided influence pooling (AGIPool), which prioritizes influential molecular substructures. The framework further employs mutual information minimization regularization to enhance information diversity during embedding fusion. Experimental results demonstrate that MolecBioNet outperforms state-of-the-art methods in DDI prediction, while ablation studies and embedding visualizations further validate the advantages of unified drug pair modeling and multi-scale knowledge integration.

Figures

Figures reproduced from arXiv: 2507.09173 by the authors.

Figure 1
Figure 1. The process of constructing a substructure-level [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed MolecBioNet framework (a, b, c) and a schematic illustration of the attention-guided [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance of MolecBioNet and its variants on the two datasets. (A) Performance for the Ryu’s dataset. (B) Perfor [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Sensitivity analysis of hyperparameters on model [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 6
Figure 6. Figure 6: Local subgraph extracted from tsBKG for Itracona [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 47 canonical work pages

  1. [1]

    Mohammad Hussain Al-Rabeah and Amir Lakizadeh. 2022. Prediction of drug- drug interaction events using graph neural networks based feature extraction. Scientific Reports 12 (Sept. 2022), 15590. doi:10.1038/s41598-022-19999-4 KDD ’25, August 3–7, 2025, Toronto, ON, Canada Mengjie Chen, Ming Zhang, & Cunquan Qu

    doi:10.1038/s41598-022-19999-4 2022

  2. [2]

    Kenza Amara, Zhitao Ying, Zitao Zhang, Zhichao Han, Yang Zhao, Yinan Shan, Ulrik Brandes, Sebastian Schemm, and Ce Zhang. 2022. GraphFramEx: Towards Systematic Evaluation of Explainability Methods for Graph Neural Networks. In The First Learning on Graphs Conference . https://openreview.net/forum?id= rGVGf1T-dK

    work page 2022

  3. [3]

    Ingolf Cascorbi. 2012. Drug Interactions—Principles, Examples and Clinical Consequences. Dtsch Arztebl International 109, 33-34 (Aug. 2012), 546–555. doi:10. 3238/arztebl.2012.0546

    arXiv 2012

  4. [4]

    Chamikara and Yi-Ping Phoebe Chen

    M.A.P. Chamikara and Yi-Ping Phoebe Chen. 2021. MedFused: A framework to discover the relationships between drug chemical functional group impacts and side effects. Computers in Biology and Medicine 133 (June 2021), 104361. doi:10.1016/j.compbiomed.2021.104361

    work page arXiv 2021

  5. [6]

    Kovács, and Albert-László Barabási

    Feixiong Cheng, István A. Kovács, and Albert-László Barabási. 2019. Network- based prediction of drug combinations. Nature Communications 10 (March 2019),

    work page 2019

  6. [7]

    Jörg Degen, Christof Wegscheid-Gerlach, Andrea Zaliani, and Matthias Rarey

    work page

  7. [8]

    Yifan Deng, Xinran Xu, Yang Qiu, Jingbo Xia, Wen Zhang, and Shichao Liu. 2020. A multimodal deep learning framework for predicting drug–drug interaction events. Bioinformatics 36, 15 (Aug. 2020), 4316–4322. doi:10.1093/bioinformatics/ btaa501

    doi:10.1093/bioinformatics/ 2020

  8. [9]

    Pieter Dewulf, Michiel Stock, and Bernard De Baets. 2021. Cold-Start Problems in Data-Driven Prediction of Drug–Drug Interaction Effects. Pharmaceuticals 14, 5 (May 2021), 429. doi:10.3390/ph14050429

    work page doi:10.3390/ph14050429 2021

Show all 71 references
  1. [10]

    Haotong Du, Quanming Yao, Juzheng Zhang, Yang Liu, and Zhen Wang. 2024. Customized Subgraph Selection and Encoding for Drug-drug Interaction Predic- tion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. https://openreview.net/forum?id=crlvDzDPgM

    2024

  2. [11]

    Vijay Prakash Dwivedi and Xavier Bresson. 2021. A Generalization of Trans- former Networks to Graphs. arXiv:2012.09699 [cs.LG] https://arxiv.org/abs/2012. 09699

    2021 arXiv

  3. [12]

    Reza Ferdousi, Reza Safdari, and Yadollah Omidi. 2017. Computational prediction of drug-drug interactions based on drugs functional similarities. Journal of Biomedical Informatics 70 (June 2017), 54–64. doi:10.1016/j.jbi.2017.04.021

    2017 doi

  4. [13]

    Weiwei Gu, Fei Gao, Xiaodan Lou, and Jiang Zhang. 2021. Discovering latent node Information by graph attention network. Scientific Reports 11 (March 2021),

    2021

  5. [14]

    Iyer, Yihong Ma, Olaf Wiest, Xiangliang Zhang, Wei Wang, Chuxu Zhang, and Nitesh V

    Zhichun Guo, Kehan Guo, Bozhao Nan, Yijun Tian, Roshni G. Iyer, Yihong Ma, Olaf Wiest, Xiangliang Zhang, Wei Wang, Chuxu Zhang, and Nitesh V. Chawla

  6. [15]

    Jukka Hakkola, Janne Hukkanen, Miia Turpeinen, and Olavi Pelkonen. 2020. Inhibition and induction of CYP enzymes in humans: an update. Archives of Toxicology 94 (Oct. 2020), 3671–3722. doi:10.1007/s00204-020-02936-7

    2020 doi

  7. [16]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, USA) (NIPS’17). Curran Associates Inc., Red Hook, NY...

    2017

  8. [17]

    Harrold and Robin M

    Marc W. Harrold and Robin M. Zavod. 2014. Basic Concepts in Medicinal Chemistry. Drug Development and Industrial Pharmacy 40, 7 (2014), 988–988. doi:10.3109/03639045.2013.789908

    2014

  9. [18]

    Changxiang He, Yuru Liu, Hao Li, Hui Zhang, Yaping Mao, Xiaofei Qin, Lele Liu, and Xuedian Zhang. 2022. Multi-type feature fusion based on graph neural network for drug-drug interaction prediction. BMC Bioinformatics 23 (June 2022),

    2022

  10. [19]

    Yue Hong, Pengyu Luo, Shuting Jin, and Xiangrong Liu. 2022. LaGAT: link-aware graph attention network for drug–drug interaction prediction. Bioinformatics 38, 24 (Dec. 2022), 5406–5412. doi:10.1093/bioinformatics/btac682

    2022 doi

  11. [20]

    Kexin Huang, Cao Xiao, Trong Hoang, Lucas Glass, and Jimeng Sun. 2020. CASTER: Predicting Drug Interactions with Chemical Substructure Representa- tion. Proceedings of the AAAI Conference on Artificial Intelligence 34, 01 (April 2020), 702–709. doi:10.1609/aaai.v34i01.5412

    2020 doi

  12. [21]

    Cao, Yixue X

    Jia Jia, Feng Zhu, Xiaohua Ma, Zhiwei W. Cao, Yixue X. Li, and Yu Zong Chen

  13. [22]

    Rezaul Karim, Michael Cochez, Joao Bosco Jares, Mamtaz Uddin, Oya Beyan, and Stefan Decker

    Md. Rezaul Karim, Michael Cochez, Joao Bosco Jares, Mamtaz Uddin, Oya Beyan, and Stefan Decker. 2019. Drug-Drug Interaction Prediction Based on Knowledge Graph Embeddings and Convolutional-LSTM Network. In Proceedings of the 10th ACM International Conference on Bioinformatics,...

    2019

  14. [23]

    Andrej Kastrin, Polonca Ferk, and Brane Leskošek. 2018. Predicting potential drug- drug interactions on topological and semantic similarity features using statistical learning. PLOS ONE 13, 5 (May 2018), 54–64. doi:10.1371/journal.pone.0196865

    2018 doi

  15. [24]

    Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, Leonid Zaslavsky, Jian Zhang, and Evan E Bolton. 2023. PubChem 2023 update. Nucleic Acids Research 51, D1 (Jan. 2023), D1373–D1380. doi:10.1093/...

    2023 doi

  16. [25]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Repre- sentations. https://openreview.net/forum?id=SJU4ayYgl

    2017

  17. [26]

    Craig Knox, Mike Wilson, Christen M Klinger, Mark Franklin, Eponine Oler, Alex Wilson, Allison Pon, Jordan Cox, Na Eun (Lucy) Chin, Seth A Straw- bridge, Marysol Garcia-Patino, Ray Kruger, Aadhavya Sivakumaran, Selena San- ford, Rahil Doshi, Nitya Khetarpal, Omolola Fatokun, D...

    2024

  18. [27]

    Longlong Li, Yipeng Zhang, Guanghui Wang, and Kelin Xia. 2024. KA-GNN: Kolmogorov-Arnold Graph Neural Networks for Molecular Property Prediction. arXiv:2410.11323 [cs.LG] https://arxiv.org/abs/2410.11323

    2024 arXiv

  19. [28]

    Zimeng Li, Shichao Zhu, Bin Shao, Xiangxiang Zeng, Tong Wang, and Tie-Yan Liu

  20. [29]

    Xuan Lin, Zhe Quan, Zhi-Jie Wang, Tengfei Ma, and Xiangxiang Zeng. 2020. KGNN: Knowledge Graph Neural Network for Drug-Drug Interaction Prediction. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, Christian Bessiere (Ed.)....

    2020 doi

  21. [30]

    Zun Liu, Xing-Nan Wang, Hui Yu, Jian-Yu Shi, and Wen-Min Dong. 2022. Predict multi-type drug–drug interactions in cold start scenario. BMC Bioinformatics 23 (Feb. 2022), 75. doi:10.1186/s12859-022-04610-4

    2022 doi

  22. [31]

    Tom Lynch and Amy Price. 2007. The effect of cytochrome P450 metabolism on drug response, interactions, and adverse effects. American Family Physician 76, 3 (Aug. 2007), 391–396

    2007

  23. [32]

    Tengfei Lyu, Jianliang Gao, Ling Tian, Zhao Li, Peng Zhang, and Ji Zhang. 2021. MDNN: A Multimodal Deep Neural Network for Predicting Drug-Drug Interac- tion Events. In Proceedings of the Thirtieth International Joint Conference on Arti- ficial Intelligence, IJCAI-21, Zhi-Hua ...

    2021 doi

  24. [33]

    Reza Bayat Mokhtari, Tina S Homayouni, Narges Baluch, Evgeniya Morgatskaya, Sushil Kumar, Bikul Das, and Herman Yeger. 2017. Combination therapy in combating cancer. Oncotarget 8, 23 (March 2017), 38022–38043. doi:10.18632/ oncotarget.16723

    2017

  25. [34]

    H. L. Morgan. 1965. The Generation of a Unique Machine Description for Chem- ical Structures-A Technique Developed at Chemical Abstracts Service. Journal of Chemical Documentation 5, 2 (May 1965), 107–113. doi:10.1021/c160017a018

    1965 doi

  26. [35]

    Briefings in Bioinformatics 24, 1 (Jan

    DSN-DDI: an accurate and generalized framework for drug–drug interac- tion prediction by dual-view representation learning. Briefings in Bioinformatics 24, 1 (Jan. 2023), bbac597. doi:10.1093/bib/bbac597

    2023 doi

  27. [36]

    Arnold K Nyamabo, Hui Yu, Zun Liu, and Jian-Yu Shi. 2022. Drug–drug interaction prediction with learnable size-adaptive molecular substructures. Briefings in Bioinformatics 23, 1 (Jan. 2022), bbab441. doi:10.1093/bib/bbab441

    2022 doi

  28. [37]

    Arnold K Nyamabo, Hui Yu, and Jian-Yu Shi. 2021. SSI–DDI: substruc- ture–substructure interactions for drug–drug interaction prediction. Briefings in Bioinformatics 22, 6 (Nov. 2021), bbab133. doi:10.1093/bib/bbab133

    2021 doi

  29. [38]

    Jae Yong Ryu, Hyun Uk Kim, and Sang Yup Lee. 2018. Deep learning improves prediction of drug–drug and drug–food interactions. Proceedings of the Na- tional Academy of Sciences 115, 18 (March 2018), E4304–E4311. doi:10.1073/pnas. 1803294115

    2018 doi

  30. [39]

    Xiaorui Su, Lun Hu, Zhuhong You, Pengwei Hu, and Bowei Zhao. 2022. Attention- based Knowledge Graph Representation Learning for Predicting Drug-drug In- teractions. Briefings in Bioinformatics 23, 3 (May 2022), bbac140. doi:10.1093/bib/ bbac140

    2022 doi

  31. [40]

    Yu, and Lun Hu

    Xiaorui Su, Pengwei Hu, Zhu-Hong You, Philip S. Yu, and Lun Hu. 2024. Dual- Channel Learning Framework for Drug-Drug Interaction Prediction via Relation- Aware Heterogeneous Graph Transformer. Proceedings of the AAAI Conference on Artificial Intelligence 38, 1 (March 2024), 24...

    2024 doi

  32. [41]

    Zhenchao Tang, Guanxing Chen, Hualin Yang, Weihe Zhong, and Calvin Yu- Chian Chen. 2024. DSIL-DDI: A Domain-Invariant Substructure Interaction Learning for Generalizable Drug–Drug Interaction Prediction. IEEE Transactions on Neural Networks and Learning Systems 35, 8 (Aug. 202...

    2024

  33. [42]

    Claude Nadeau and Yoshua Bengio. 1999. Inference for the generalization error. In Proceedings of the 13th International Conference on Neural Information Processing Systems (Denver, CO) (NIPS ’99). MIT Press, Cambridge, MA, USA, 307–313

    1999

  34. [43]

    Tatonetti, Patrick P

    Nicholas P. Tatonetti, Patrick P. Ye, Roxana Daneshjou, and Russ B. Altman. 2012. Data-Driven Prediction of Drug Effects and Interactions. Science Translational Medicine 4 (March 2012), 125ra31. doi:10.1126/scitranslmed.3003377

    2012 doi

  35. [44]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. InInternational Confer- ence on Learning Representations . https://openreview.net/forum?id=rJXMpikCZ

    2018

  36. [45]

    Santiago Vilar, Rave Harpaz, Eugenio Uriarte, Lourdes Santana, Raul Rabadan, and Carol Friedman. 2012. Drug—drug interaction through molecular structure similarity analysis. Journal of the American Medical Informatics Association 19, 6 (Nov. 2012), 1066–1074. doi:10.1136/amiaj...

    2012 doi

  37. [46]

    Santiago Vilar, Eugenio Uriarte, Lourdes Santana, Tal Lorberbaum, George Hripc- sak, Carol Friedman, and Nicholas P Tatonetti. 2014. Similarity-based modeling in large-scale prediction of drug-drug interactions. Nature Protocols 9 (Aug. 2014), 2147–2163. doi:10.1038/nprot.2014.151

    2014 doi

  38. [47]

    Thanh Hoa Vo, Ngan Thi Kim Nguyen, Quang Hien Kha, and Nguyen Quoc Khanh Le. 2022. On the road to explainable AI in drug-drug interactions prediction: A systematic review. Computational and Structural Biotechnology Journal 20 (April 2022), 2112–2123. doi:10.1016/j.csbj.2022.04.021

    2022 doi

  39. [48]

    Yaqing Wang, Zaifei Yang, and Quanming Yao. 2024. Accurate and interpretable drug-drug interaction prediction enabled by knowledge subgraph learning. Com- munications Medicine 4 (March 2024), 59. doi:10.1038/s43856-024-00486-y

    2024 doi

  40. [49]

    Farhan Tanvir, Muhammad Ifte Khairul Islam, and Esra Akbas. 2021. Predicting Drug-Drug Interactions Using Meta-path Based Similarities. In 2021 IEEE Confer- ence on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB). 1–8. doi:10.1109/CIBCB49929.2021.9562802

    2021

  41. [50]

    Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. 2016. A Discriminative Feature Learning Approach for Deep Face Recognition. In Computer Vision – ECCV 2016, Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling (Eds.). Springer International Publishing, Cham, 499–515

    2016

  42. [51]

    Steven Whitebread, Jacques Hamon, Dejan Bojanic, and Laszlo Urban. 2005. Keynote review: In vitro safety pharmacology profiling: an essential tool for successful drug development. Drug Discovery Today 10, 21 (Nov. 2005), 1421–

    2005

  43. [52]

    Grossman

    Tian Xie and Jeffrey C. Grossman. 2018. Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties. Phys. Rev. Lett. 120 (Apr 2018), 145301. Issue 14. doi:10.1103/PhysRevLett.120. 145301

    2018 doi

  44. [53]

    Nuo Xu, Pinghui Wang, Long Chen, Jing Tao, and Junzhou Zhao. 2019. MR-GNN: Multi-Resolution and Dual Graph Neural Network for Predicting Structured Entity Interactions. In Proceedings of the Twenty-Eighth International Joint Con- ference on Artificial Intelligence, IJCAI-19 . ...

    2019 doi

  45. [54]

    Hui Yu, Kui-Tao Mao, Jian-Yu Shi, Hua Huang, Zhi Chen, Kai Dong, and Siu-Ming Yiu. 2018. Predicting and understanding comprehensive drug-drug interactions via semi-nonnegative matrix factorization. BMC Systems Biology 12 (April 2018),

    2018

  46. [55]

    Hui Yu, Jing Wang, Shi-Yu Zhao, Omayo Silver, Zun Liu, JingTao Yao, and Jian-Yu Shi. 2024. GGI-DDI: Identification for key molecular substructures by granule learning to interpret predicted drug–drug interactions. Expert Systems with Applications 240 (April 2024), 122500. doi:...

    2024

  47. [56]

    Zhi-Yu Wang, Meng Chen, Ling-Ling Zhu, Lu-Shan Yu, Su Zeng, Mei-Xiang Xiang, and Quan Zhou. 2015. Pharmacokinetic drug interactions with clopidogrel: updated review and risk management in combination therapy. Therapeutics and Clinical Risk Management 11 (March 2015), 449–67. d...

    2015 doi

  48. [57]

    Hao Yuan, Haiyang Yu, Shurui Gui, and Shuiwang Ji. 2023. Explainability in Graph Neural Networks: A Taxonomic Survey.IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 5 (2023), 5782–5799. doi:10.1109/TPAMI.2022.3204236

    2023

  49. [58]

    Jing Zhang, Deng-Ping Fan, Yuchao Dai, Xin Yu, Yiran Zhong, Nick Barnes, and Ling Shao. 2021. RGB-D Saliency Detection via Cascaded Mutual Informa- tion Minimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 4338–4347

    2021

  50. [59]

    Muhan Zhang and Yixin Chen. 2018. Link prediction based on graph neural net- works. In Proceedings of the 32nd International Conference on Neural Information Processing Systems (Montréal, Canada) (NIPS’18). Curran Associates Inc., Red Hook, NY, USA, 5171–5181

    2018

  51. [60]

    Yongqi Zhang, Quanming Yao, Ling Yue, Xian Wu, Ziheng Zhang, Zhenxi Lin, and Yefeng Zheng. 2023. Emerging drug interaction prediction enabled by a flow- based graph neural network with biomedical network. Nature Computational Science 3 (Dec. 2023), 1023–1033. doi:10.1038/s4358...

    2023 doi

  52. [61]

    Yi Zhong, Gaozheng Li, Ji Yang, Houbing Zheng, Yongqiang Yu, Jiheng Zhang, Heng Luo, Biao Wang, and Zuquan Weng. 2024. Learning motif-based graphs for drug–drug interaction prediction via local–global self-attention. Nature Machine Intelligence 6 (Sept. 2024), 1094–1105. doi:1...

    2024 doi

  53. [62]

    Marinka Zitnik, Monica Agrawal, and Jure Leskovec. 2018. Modeling polyphar- macy side effects with graph convolutional networks. Bioinformatics 34, 13 (July 2018), i457–i466. doi:10.1093/bioinformatics/bty294 A Molecular Graph Representation via GCN To compute the graph-level ...

    2018 doi

  54. [63]

    doi:10.1186/s12918-018-0532-7

    doi

  55. [65]

    Yue Yu, Kexin Huang, Chao Zhang, Lucas M Glass, Jimeng Sun, and Cao Xiao

  56. [224]

    doi:10.1186/s12859-022-04763-2

    doi

  57. [1197]

    doi:10.1038/s41467-019-09186-x

    doi

  58. [1433]

    doi:10.1016/S1359-6446(05)03632-9

    doi

  59. [2008]

    ChemMedChem 3, 10 (Oct

    On the Art of Compiling and Using ‘Drug-Like’ Chemical Fragment Spaces. ChemMedChem 3, 10 (Oct. 2008), 1503–1507. doi:10.1002/cmdc.200800178

    2008 doi

  60. [2009]

    Nature Reviews Drug Discovery 8 (Feb

    Mechanisms of drug combinations: interaction and network perspectives. Nature Reviews Drug Discovery 8 (Feb. 2009), 111–128. doi:10.1038/nrd2683

    2009 doi

  61. [2021]

    Bioinformatics 37, 18 (Sept

    SumGNN: multi-typed drug interaction prediction via efficient knowledge graph summarization. Bioinformatics 37, 18 (Sept. 2021), 2988–2995. doi:10.1093/ bioinformatics/btab207

    2021

  62. [2023]

    In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23 , Edith Elkind (Ed.)

    Graph-based Molecular Representation Learning. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23 , Edith Elkind (Ed.). International Joint Conferences on Artificial Intelligence Organization, 6638–6646. doi:10.24963/ijcai.2...

    doi

  63. [6967]

    doi:10.1038/s41598-021-85826-x

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.