Deep Learning Framework MDM-DDI Improves Drug-Drug Interaction Predictions Across Benchmarks

Administering multiple medications concurrently represents a standard therapeutic approach for managing complex diseases, yet simultaneous drug consumption significantly elevates the risk of severe adverse reactions caused by drug-drug interactions. These interactions, spanning synergistic, antagonistic, or disruptive physiological effects, range in severity from reduced treatment efficacy to fatal emergencies, posing a substantial clinical challenge for elderly populations where twenty to forty percent of individuals taking multiple prescriptions experience interaction events. In a study published in Frontiers in Pharmacology, researchers Yiqun Wang, Shilong Song, and Yufei Li from Gansu University of Political Science and Law introduced MDM-DDI, a deep learning architecture designed to improve computational interaction prediction. The system characterizes drug molecules across four dimensions: one-dimensional chemical sequences, two-dimensional topological graphs, three-dimensional spatial conformations, and pharmacological textual knowledge. Experiments across benchmark datasets demonstrate that MDM-DDI achieves consistent performance gains over established baseline methods while improving predictive stability when evaluating previously unseen pharmaceutical compounds.

Traditionally, identifying drug interactions has depended on empirical clinical trials and laboratory in vitro experiments, methods that remain costly, labor-intensive, and incapable of screening vast combinatorial libraries of chemical pairs. To overcome these constraints, computational biology has increasingly embraced data-driven prediction models. Early machine learning algorithms relied heavily on drug similarity hypotheses, assuming that compounds structurally resembling a known interacting drug would exhibit comparable physiological behaviors. However, these foundational models, which often paired multi-attribute Tanimoto similarity metrics with logistic regression or random forest classifiers, suffered from critical limitations: feature selection depended on subjective domain expertise, introducing substantial representation bias and feature noise that degraded predictive accuracy. Although subsequent deep learning models autonomously extracted deeper representations from molecular sequences or graph topologies using recurrent units and graph attention layers, existing architectures still struggled to adequately characterize the multifaceted physicochemical properties and intricate spatial conformations of complex pharmaceutical molecules, leaving critical structural data underutilized.

In recent years, large language models pre-trained on extensive biomedical literature offered an alternative strategy by providing rich pharmacological knowledge concerning drug indications and biochemical mechanisms. Computational approaches such as BioBERT-DDI and graph-coupled language models demonstrated that incorporating textual embeddings could enhance predictive capability, especially in low-resource settings. Nevertheless, these textual representations introduce significant drawbacks. Embeddings generated by large language models frequently carry substantial irrelevant information, semantic noise, and descriptive redundancy that fail to reflect atomic-level interaction mechanisms, thereby limiting prediction reliability. Addressing these issues, the MDM-DDI framework targets three persistent shortcomings in computer-aided pharmacology: the incomplete extraction of multi-dimensional structural and stereochemical data, the absence of bidirectional associative modeling between paired pharmaceutical agents, and the persistent noise embedded within uncalibrated textual features extracted from general biomedical language models. By integrating multi-modal structural encoders with dedicated denoising mechanisms, the framework attempts to improve both overall accuracy and generalization across diverse interaction scenarios.

To represent drug molecules comprehensively, MDM-DDI integrates four complementary input dimensions covering chemical, structural, geometric, and textual domain properties. The one-dimensional sequence utilizes the Simplified Molecular Input Line Entry System string to record atomic composition and bonding order, reflecting the basic chemical skeleton. The two-dimensional molecular graph, constructed via RDKit, treats atoms as nodes containing intrinsic elemental features like atomic number and electronegativity, while chemical bonds serve as edges carrying explicit bond classifications. For the three-dimensional geometric representation, the authors applied RDKit’s experimental torsion-angle distance geometry algorithm followed by local energy minimization under the Merck Molecular Force Field, selecting the lowest-energy conformation to record atomic coordinates, inter-atomic Euclidean distances, and stereochemical orientations. Finally, pharmacological prior knowledge was retrieved from official drug monographs in the DrugBank database, covering mechanisms of action and clinical indications. Sentences explicitly mentioning other drug names or specific interaction events were systematically excluded to prevent label leakage.

To process these distinct modalities, the framework incorporates four specialized neural feature extraction encoders. The one-dimensional sequence encoder adopts a hybrid architecture combining a bidirectional long short-term memory network with a Transformer encoder, capturing localized chemical substructures alongside long-range sequential context. The two-dimensional graph encoder applies a Graph Isomorphism Network with Edge features, which linearly transforms bond attributes to align them with atomic node features, updating molecular representations layer by layer. For three-dimensional conformations, the researchers implemented an Equivariant Graph Neural Network, which updates spatial coordinates and atomic attributes while preserving rotational, translational, and reflective equivariance. Meanwhile, the text feature encoder utilizes BioGPT, a generative language model trained on biomedical corpora. DrugBank description texts are processed through BioGPT, where mean pooling across valid hidden-layer tokens yields a fixed-dimensional semantic vector that is projected into an initial text embedding, transforming unstructured literature into standardized machine-readable features for downstream processing.

Because drug interactions represent mutual, often asymmetrical biochemical relationships between paired molecules, MDM-DDI incorporates a bidirectional drug interaction module rather than relying on simple feature concatenation. Operating independently across sequence, graph, and geometric modalities, bidirectional cross-attention layers enable atomic representations of each drug to integrate information from the corresponding partner molecule. These interaction matrices are compressed into fixed-length vectors using multi-head attention pooling, with gating mechanisms dynamically adjusting directional weights. To counter textual noise from BioGPT embeddings, the authors designed a Memory-Augmented Network inspired by external memory architectures. During inference, a query vector constructed from combined multimodal interaction features and drug text embeddings reads the memory slots via scaled dot-product attention, producing a denoised knowledge vector. A momentum-based update mechanism dynamically refines the memory slots during training without gradient backpropagation.

For final classification, the three multimodal interaction vectors and the denoised memory read vector are stacked and merged through multi-head self-attention pooling, adaptively learning contribution weights across modalities. The fused representation passes into a multi-layer perceptron with residual connections and layer normalization to predict logits across interaction types. The authors implemented the model using PyTorch and Python 3.8 on an NVIDIA RTX 3090 GPU, training with AdamW under a batch size of 64 and a memory update rate of 0.01. Testing utilized two public benchmarks: DrugBank, containing 191,878 records across 1,706 drugs and 86 interaction types evaluated as a multi-class task, and TWOSIDES, comprising 4,576,287 records across 645 drugs and 963 adverse reaction types evaluated as a multi-label task. Experiments employed stratified five-fold cross-validation repeated three times. Performance metrics included Accuracy, Macro-AUC, Macro-AUPR, and Macro-F1 for multi-class classification, alongside Micro-AUC and Micro-AUPR for multi-label evaluation.

Experimental evaluations demonstrated that MDM-DDI achieved the highest performance among all tested models on both benchmark datasets. In the DrugBank multi-class evaluation, MDM-DDI reached an Accuracy of 0.968, Macro-AUC of 0.991, Macro-AUPR of 0.960, and Macro-F1 of 0.937, consistently outperforming baselines including Deep DDI, SSI-DDI, SUM-GNN, MUFFIN, and LMF-DDI. Compared to the leading baseline LMF-DDI, MDM-DDI improved Accuracy by 0.6 percentage points, Macro-AUPR by 0.5 percentage points, and Macro-F1 by 0.5 percentage points. Paired t-tests confirmed that these performance gains were statistically significant, yielding p-values below 0.05. On the TWOSIDES multi-label dataset, MDM-DDI achieved a Micro-AUC of 0.913 and a Micro-AUPR of 0.696, surpassing MUFFIN by 0.4 and 1.7 percentage points, respectively. The researchers emphasized that precision-recall area metrics like Macro-AUPR and Micro-AUPR are exceptionally sensitive to class imbalance, indicating that MDM-DDI substantially improves classification accuracy for rare, low-frequency interaction categories that typically challenge conventional models.

To evaluate generalization to newly developed medications, the team conducted cold-start experiments on DrugBank with twenty percent unseen drugs across five random seeds. Under setting S1, involving one unseen drug per pair, MDM-DDI achieved an Accuracy of 0.767 and Macro-F1 of 0.706, exceeding LMF-DDI’s 0.756 and 0.690. In the more demanding S2 setting, where both drugs were unseen, MDM-DDI attained an Accuracy of 0.593 and Macro-F1 of 0.495, outperforming LMF-DDI’s 0.580 and 0.484, respectively. Ablation studies confirmed the necessity of each architectural component: removing two-dimensional graph features or three-dimensional conformations produced the largest performance declines, reducing Macro-F1 by 1.5 and 1.3 percentage points and lowering Macro-AUPR to 0.943. Removing the bidirectional module reduced Macro-F1 to 0.927. Furthermore, removing the memory network while retaining raw text yielded a Macro-F1 of 0.931, while replacing it with a multilayer perceptron reached 0.932, underscoring the prototype network’s specific role in filtering textual noise.

The researchers validated practical utility through case studies on two novel FDA-approved medications from 2023 absent from training data: Fezolinetant and Elacestrant. Evaluating 3,412 candidate pairings against training compounds, the model’s highest-confidence predictions—such as Fezolinetant interacting with Fluvoxamine, Caffeine, and Warfarin, alongside Elacestrant pairing with Ketoconazole, Fulvestrant, and Palbociclib—were all confirmed by existing biomedical literature and clinical labels. Interpretability analyses on the Fezolinetant-Fluvoxamine pair showed that attention mechanisms concentrated on Fezolinetant’s 4-fluorophenyl metabolic site and Fluvoxamine’s CYP1A2-inhibiting trifluoromethyl and methoxy groups, mirroring established enzymatic pathways. However, the authors noted that attention weights provide only auxiliary observations rather than definitive causal proofs, observing minor attention dispersion across non-pharmacophore regions due to pair-level data supervision. They acknowledged key study limitations, including the exclusion of target protein microenvironments, the absence of non-interacting negative samples, and the ongoing need for knowledge calibration mechanisms to prevent large language model hallucinations in automated predictive pharmacology.

Subject of Research: Deep learning framework for predicting multi-dimensional drug-drug interactions using memory-augmented textual denoising

Article Title: MDM-DDI: a drug–drug interaction prediction framework integrating multi-dimensional molecular features and memory-augmented networks

Article References: Wang, Y., Song, S., & Li, Y. (2026). MDM-DDI: a drug–drug interaction prediction framework integrating multi-dimensional molecular features and memory-augmented networks. Frontiers in Pharmacology, 17, Article 1951995. https://doi.org/10.3389/fphar.2026.1951995

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1951995

Keywords: drug-drug interactions, MDM-DDI framework, pharmacological prior knowledge, memory-augmented network, BioGPT, graph isomorphism network, equivariant graph neural network, molecular representation, DrugBank dataset, TWOSIDES dataset, cold-start generalization, CYP1A2 metabolism, fezolinetant, elacestrant, multi-modal deep learning, cross-attention mechanism