MEGH-DDI: MolT5-driven edge-aware graph learning with Hop2Token representation for drug–drug interaction prediction
- 1School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China.
- 2School of Artificial Intelligence, Hebei University of Technology, Tianjin, China.
Abstract
Drug-drug interactions (DDIs) are critical in pharmacology, as synergistic or antagonistic effects between co-administered drugs can enhance therapeutic outcomes or cause toxicity, posing major challenges for drug safety assessment and new drug development. Current molecular representation methods often fail to capture dynamic substructure relationships or semantic interactions among functional groups. To overcome these limitations, this study proposes MEGH-DDI, a multimodal framework that integrates complementary molecular representation learning with graph-structure dynamic perception. The framework employs a dual-channel semantic module that combines Mol2vec-based substructure representations with MolT5 chemical language embeddings, while SHAP-based dynamic feature selection reduces redundancy. For fine-grained structural modeling, we introduce an edge-aware GCNII equipped with bond-level attention to distinguish different types of chemical bonds. Furthermore, the Hop2Token module converts multi-hop neighborhoods into token sequences, enabling Transformer-based modeling of high-order substructure dependencies. Experimental results demonstrate that MEGH-DDI achieves competitive performance across both transductive and inductive settings. On the DrugBank dataset, MEGH-DDI attains an accuracy of 98.36% in the transductive task, outperforming earlier baselines and surpassing recent state-of-the-art methods, with both AUC and AP exceeding 99.7%. On the ChChMiner dataset, MEGH-DDI achieves an F1 score of 99.05%, improving over the best existing method by 2.74 percentage points. In the inductive settings, MEGH-DDI achieves competitive performance across most evaluation metrics on both datasets, showing advantages over competing methods on several key metrics. These results demonstrate the effectiveness of MEGH-DDI in predicting DDI involving previously unseen drugs and its potential to generalize to complex DDI prediction scenarios. To enhance interpretability, MEGH-DDI incorporates a three-level mechanism spanning atoms, chemical bonds, and topological paths, highlighting key atoms and bond contributions. Overall, MEGH-DDI provides an efficient, accurate, and interpretable solution for drug safety assessment and new drug discovery. Code and datasets are available at https://github.com/Lilyjoys/MEGH-DDI, and a free web platform is provided at http://www.edgh-ddi.com. Author summary: Drug–drug interactions (DDIs) are critically important in pharmacology, as co-administered drugs can produce synergistic or antagonistic effects that significantly affect therapeutic outcomes and patient safety. However, existing computational methods often struggle to capture the complex semantic and structural dependencies among molecular substructures, particularly in scenarios involving unseen drugs or small-sample datasets. To overcome these limitations, we propose MEGH-DDI, a multimodal and interpretable framework for DDI prediction. MEGH-DDI integrates molecular semantic information and graph structural features through a dual-channel design: at the semantic level, it combines Mol2vec substructure embeddings with MolT5 chemical language representations and employs SHAP-based dynamic feature selection to remove redundant features; at the structural level, it introduces an edge-aware GCNII with bond-level attention and leverages the Hop2Token module to serialize multi-hop neighborhoods, enabling Transformer-based modeling of high-order structural dependencies.
Related
- HierDTA: Hierarchical Structure-Aware Representation Learning for Drug–Target Affinity Prediction
- P2MAT: A machine learning (ML) driven software for Property Prediction of MATerial
- Sustainable hemocompatibility prediction of electrospun nanomaterials using SEM-driven CNN ensemble learning aligned with ISO 10993
- Machine Learning-Driven Comparative Prediction of Compressive Strength in True Geopolymer and Hybrid Alkali-Activated Cement Systems
- Machine Learning-Driven Prediction and Optimization of Thermoelectric Transport Properties in Ce-Doped CaMnO3 Ceramics
BibTeX
@article{zhou2026meghddi,
title = {{MEGH-DDI: MolT5-driven edge-aware graph learning with Hop2Token representation for drug–drug interaction prediction}},
author = {Li Zhou and Yun Zuo and Wenying He and Zhaohong Deng},
journal = {PLOS Computational Biology},
year = {2026},
month = oct,
doi = {10.1371/journal.pcbi.1014393},
url = {https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014393},
note = {CC BY},
}
RIS
TY - JOUR
TI - MEGH-DDI: MolT5-driven edge-aware graph learning with Hop2Token representation for drug–drug interaction prediction
AU - Li Zhou
AU - Yun Zuo
AU - Wenying He
AU - Zhaohong Deng
PY - 2026
DA - 2026/10/14
JO - PLOS Computational Biology
DO - 10.1371/journal.pcbi.1014393
UR - https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014393
AB - Drug-drug interactions (DDIs) are critical in pharmacology, as synergistic or antagonistic effects between co-administered drugs can enhance therapeutic outcomes or cause toxicity, posing major challenges for drug safety assessment and new drug development. Current molecular representation methods often fail to capture dynamic substructure relationships or semantic interactions among functional groups. To overcome these limitations, this study proposes MEGH-DDI, a multimodal framework that integrates complementary molecular representation learning with graph-structure dynamic perception. The framework employs a dual-channel semantic module that combines Mol2vec-based substructure representations with MolT5 chemical language embeddings, while SHAP-based dynamic feature selection reduces redundancy. For fine-grained structural modeling, we introduce an edge-aware GCNII equipped with bond-level attention to distinguish different types of chemical bonds. Furthermore, the Hop2Token module converts multi-hop neighborhoods into token sequences, enabling Transformer-based modeling of high-order substructure dependencies. Experimental results demonstrate that MEGH-DDI achieves competitive performance across both transductive and inductive settings. On the DrugBank dataset, MEGH-DDI attains an accuracy of 98.36% in the transductive task, outperforming earlier baselines and surpassing recent state-of-the-art methods, with both AUC and AP exceeding 99.7%. On the ChChMiner dataset, MEGH-DDI achieves an F1 score of 99.05%, improving over the best existing method by 2.74 percentage points. In the inductive settings, MEGH-DDI achieves competitive performance across most evaluation metrics on both datasets, showing advantages over competing methods on several key metrics. These results demonstrate the effectiveness of MEGH-DDI in predicting DDI involving previously unseen drugs and its potential to generalize to complex DDI prediction scenarios. To enhance interpretability, MEGH-DDI incorporates a three-level mechanism spanning atoms, chemical bonds, and topological paths, highlighting key atoms and bond contributions. Overall, MEGH-DDI provides an efficient, accurate, and interpretable solution for drug safety assessment and new drug discovery. Code and datasets are available at https://github.com/Lilyjoys/MEGH-DDI, and a free web platform is provided at http://www.edgh-ddi.com. Author summary: Drug–drug interactions (DDIs) are critically important in pharmacology, as co-administered drugs can produce synergistic or antagonistic effects that significantly affect therapeutic outcomes and patient safety. However, existing computational methods often struggle to capture the complex semantic and structural dependencies among molecular substructures, particularly in scenarios involving unseen drugs or small-sample datasets. To overcome these limitations, we propose MEGH-DDI, a multimodal and interpretable framework for DDI prediction. MEGH-DDI integrates molecular semantic information and graph structural features through a dual-channel design: at the semantic level, it combines Mol2vec substructure embeddings with MolT5 chemical language representations and employs SHAP-based dynamic feature selection to remove redundant features; at the structural level, it introduces an edge-aware GCNII with bond-level attention and leverages the Hop2Token module to serialize multi-hop neighborhoods, enabling Transformer-based modeling of high-order structural dependencies.
SN - 1553-7358
SN - 1553-734X
ER -