MDM-DDI: a drug–drug interaction prediction framework integrating multi-dimensional molecular features and memory-augmented networks
Abstract
When patients take multiple medications concurrently, drug–drug interactions (DDIs) can induce severe adverse reactions, making accurate DDI prediction critical for ensuring medication safety. Existing deep learning approaches often struggle to comprehensively characterize the complex molecular structures and physicochemical properties of drug molecules. Meanwhile, related studies that directly incorporate pharmacological knowledge from large language models tend to introduce redundant noise, thereby limiting predictive performance. To address these issues, we propose MDM-DDI, a framework that integrates four dimensions of drug molecular features: one-dimensional sequences, two-dimensional topologies, three-dimensional conformations, and pharmacological prior knowledge. A bidirectional drug interaction module is designed to capture latent associations between drug pairs, and a memory network mechanism is innovatively introduced to filter and denoise the pharmacological priors generated by large language models. Extensive experiments on benchmark datasets including DrugBank demonstrate that MDM-DDI achieves small but consistent improvements over the evaluated baseline methods across multiple evaluation metrics, and also shows improvements over the evaluated baseline methods under cold-start scenarios. Case studies and interpretability analyses further provide preliminary evidence for the rationality of the model predictions.
Related
- A Structure-Based hERG Pose Selection Workflow Integrating Induced-Fit Docking, Shape Similarity, Interaction Fingerprints, and Molecular Dynamics
- Natural products targeting mitochondrial homeostasis: a four-dimensional framework integrating dynamics, mitophagy, pharmacokinetics, and translational safety
- Three-dimensional multimodal fractal dynamics integrating spatial metabolism, cytoarchitecture, and perfusion for preoperative prediction of microvascular invasion in colorectal liver metastases
- Molecular Biomarker Testing in Kidney Cancer: Integrating Diagnosis, Prediction, and Emerging Renal Tumor Entities
- PRInter: an inductive multimodal heterogeneous network embedding framework with contrastive learning for lncRNA-protein interaction prediction
BibTeX
@article{wang2026mdmddi,
title = {{MDM-DDI: a drug–drug interaction prediction framework integrating multi-dimensional molecular features and memory-augmented networks}},
author = {Yiqun Wang and Shilong Song and Yufei Li},
journal = {Frontiers in Pharmacology},
year = {2026},
month = nov,
doi = {10.3389/fphar.2026.1951995},
url = {https://www.frontiersin.org/articles/10.3389/fphar.2026.1951995},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - MDM-DDI: a drug–drug interaction prediction framework integrating multi-dimensional molecular features and memory-augmented networks
AU - Yiqun Wang
AU - Shilong Song
AU - Yufei Li
PY - 2026
DA - 2026/11/09
JO - Frontiers in Pharmacology
DO - 10.3389/fphar.2026.1951995
UR - https://www.frontiersin.org/articles/10.3389/fphar.2026.1951995
AB - When patients take multiple medications concurrently, drug–drug interactions (DDIs) can induce severe adverse reactions, making accurate DDI prediction critical for ensuring medication safety. Existing deep learning approaches often struggle to comprehensively characterize the complex molecular structures and physicochemical properties of drug molecules. Meanwhile, related studies that directly incorporate pharmacological knowledge from large language models tend to introduce redundant noise, thereby limiting predictive performance. To address these issues, we propose MDM-DDI, a framework that integrates four dimensions of drug molecular features: one-dimensional sequences, two-dimensional topologies, three-dimensional conformations, and pharmacological prior knowledge. A bidirectional drug interaction module is designed to capture latent associations between drug pairs, and a memory network mechanism is innovatively introduced to filter and denoise the pharmacological priors generated by large language models. Extensive experiments on benchmark datasets including DrugBank demonstrate that MDM-DDI achieves small but consistent improvements over the evaluated baseline methods across multiple evaluation metrics, and also shows improvements over the evaluated baseline methods under cold-start scenarios. Case studies and interpretability analyses further provide preliminary evidence for the rationality of the model predictions.
SN - 1663-9812
ER -