AI Model Reads Chemical Language and Graph Structure to Predict Dangerous Drug Combinations

When patients take multiple medications at once, the drugs do not always stay in their own chemical lanes. Some combinations amplify each other’s therapeutic effects, while others collide in the body to produce toxicity that neither drug would cause alone. These drug–drug interactions, or DDIs, remain one of the most stubborn challenges in pharmacology, and identifying them experimentally is slow and expensive. A new computational framework called MEGH-DDI, described in PLOS Computational Biology by Li Zhou, Yun Zuo, Wenying He and Zhaohong Deng, now offers a way to predict these interactions with remarkable accuracy by teaching artificial intelligence to read molecules the way language models read sentences.

The core problem the researchers set out to solve lies in how existing algorithms represent drug molecules. Most current methods treat a molecule as a static picture of atoms and bonds, which means they often miss how substructures relate to each other dynamically or how the semantic meaning of one functional group changes the interpretation of another. In a drug interaction context, those overlooked relationships can be exactly the ones that matter, because the biological effect of a compound frequently emerges from the interplay between its chemical components rather than from any single fragment.

MEGH-DDI attacks this limitation with a multimodal design that pairs two complementary ways of understanding a molecule. The first is a dual-channel semantic module that combines Mol2vec, a technique that converts molecular substructures into numerical vectors, with MolT5, a chemical language model adapted from the text-processing architectures that power modern natural language systems. By running both channels in parallel, the framework captures fine-grained substructure information alongside broader chemical semantics, much as a reader might combine knowledge of individual words with an understanding of the sentence they appear in.

Because merging two representation channels can introduce redundant or noisy features, the team added a dynamic feature selection step based on SHAP, a widely used interpretability technique that estimates how much each feature contributes to a prediction. Features that add little signal are down-weighted or removed, which keeps the model focused on the molecular details that genuinely drive interactions. This pruning step is not merely cosmetic; by reducing redundancy it helps the downstream network learn cleaner patterns and improves the model’s ability to generalize beyond the drugs it has already seen.

The second pillar of the framework is structural. An edge-aware graph convolutional network, built on the GCNII architecture, treats each molecule as a graph in which atoms are nodes and chemical bonds are edges. What makes this version distinctive is its bond-level attention, which allows the network to distinguish between different types of chemical bonds rather than treating every connection identically. A single bond and an aromatic bond carry very different chemical implications, and the attention mechanism lets the model weigh those differences explicitly as information propagates through the molecular graph.

Perhaps the most inventive component is the Hop2Token module. Graph neural networks typically aggregate information from immediate neighbors, but important chemical relationships can span several bonds. Hop2Token converts multi-hop neighborhoods, meaning the atoms reachable within two, three or more steps across the molecular graph, into token sequences. These sequences can then be processed by a Transformer, the same architecture underlying modern large language models, allowing the framework to model high-order dependencies between distant substructures that conventional message-passing networks would struggle to capture.

The performance numbers reported in the study are striking. On the widely used DrugBank dataset, MEGH-DDI reached an accuracy of 98.36 percent in the transductive setting, where the model is evaluated on drug pairs it has encountered during training. Both the area under the ROC curve and average precision exceeded 99.7 percent, placing the framework ahead of earlier baselines and recent state-of-the-art methods. On the ChChMiner dataset, the model achieved an F1 score of 99.05 percent, an improvement of 2.74 percentage points over the best previously published approach.

Transductive results alone would be impressive but limited, since real-world drug safety questions usually involve newly approved compounds that no model has seen before. The researchers therefore also tested MEGH-DDI in inductive settings, where interactions involving previously unseen drugs must be predicted. The framework remained competitive across most evaluation metrics on both datasets and outperformed competing methods on several key measures, suggesting that its combination of semantic and structural learning genuinely generalizes rather than memorizing training pairs.

Accuracy alone is rarely enough in a medical context, where clinicians and regulators need to understand why a model flags a particular combination as risky. To address this, MEGH-DDI incorporates a three-level interpretability mechanism that spans atoms, chemical bonds and topological paths through the molecular graph. This design highlights which atoms and bonds contributed most to a given prediction, giving researchers a chemically meaningful explanation rather than an opaque score. Such transparency could make the tool more trustworthy for drug safety assessment and for guiding the design of new compounds with fewer interaction liabilities.

The team has made the framework broadly accessible, releasing code and datasets through a public GitHub repository and offering a free web platform where researchers can run predictions without installing anything. As polypharmacy becomes increasingly common in aging populations and combination therapies grow central to treating complex diseases, tools like MEGH-DDI point toward a future in which dangerous drug interactions can be flagged computationally before they ever reach a patient, accelerating both drug safety evaluation and the search for safer new medicines.

Subject of Research: Prediction of drug–drug interactions using multimodal molecular representation learning with graph neural networks and chemical language models

Article Title: MEGH-DDI: MolT5-driven edge-aware graph learning with Hop2Token representation for drug–drug interaction prediction

Article References: Zhou, L., Zuo, Y., He, W., & Deng, Z. (2026). MEGH-DDI: MolT5-driven edge-aware graph learning with Hop2Token representation for drug–drug interaction prediction. PLOS Computational Biology, 22(10), e1014393. https://doi.org/10.1371/journal.pcbi.1014393

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014393

Keywords: drug–drug interactions, MEGH-DDI, MolT5, graph neural networks, Transformer, molecular representation, pharmacology, drug safety, DrugBank, ChChMiner, SHAP interpretability, computational biology