Accurate prediction of drug–drug interactions remains essential for patient safety and therapeutic optimisation. Structure-based graph neural networks have shown strong performance in this field, yet many approaches treat molecular substructures uniformly and do not explicitly incorporate established medicinal chemistry knowledge. A recently described framework, FG-DDI, introduces functional groups as chemically meaningful units within molecular graphs derived from SMILES representations. By embedding functional-group enrichment signals into a dual-view message-passing architecture, the model aims to improve predictive robustness while preserving interpretability at substructure level. The approach is evaluated under both transductive and inductive scenarios, distinguishing between settings where all drugs are seen during training and those involving unseen compounds. Experiments on two established datasets assess classification performance and explore how functional-group priors influence model behaviour across interaction types.
Embedding Functional Group Knowledge in Graph Learning
FG-DDI represents each drug as a molecular graph with atoms as nodes and bonds as edges. Functional groups are detected using SMARTS-based pattern matching from a predefined list of substructures selected for pharmacological relevance and prevalence in approved drugs. The framework considers several dozen functional groups, organised into broader chemical categories that include heterocyclic systems, carboxylic derivatives, alcohols, carbonyls, nitrogen-containing groups, ethers and specialised motifs such as nitro, nitrile, phosphate, sulfonamide, guanidine and urea.
Must Read: Federated LLMs Advance Adverse Drug Reaction Prediction
To manage overlapping patterns, detection follows a hierarchical strategy in which more complex and specific motifs are prioritised before simpler groups. This ordering is intended to preserve chemically meaningful distinctions. Functional-group enrichment scores are then calculated from observed co-occurrence patterns in known interaction data, comparing the likelihood of interaction for a given pair of groups against the dataset baseline. These scores are incorporated into both intra-molecular and inter-molecular message passing through learnable gates, allowing the model to retain flexibility while leveraging curated priors.
The architecture builds on a dual-view design that separates reasoning within individual drugs from reasoning across drug pairs. In different evaluation settings, additive or multiplicative enhancement strategies are applied to integrate enrichment signals. Hyperparameters such as attention blocks, embedding dimensions and optimisation settings follow a consistent configuration across experiments.
Evaluation Across Datasets and Generalisation Settings
The framework is evaluated on DrugBank and TwoSides, two widely used benchmarks for drug–drug interaction prediction. DrugBank contains interaction triplets covering more than one thousand drugs and several dozen interaction types. TwoSides includes a substantially larger number of interaction triplets and a broad spectrum of adverse event categories derived from reporting data, with filtering applied to remove rare interaction types.
Experiments follow standard training, validation and test splits with cross-validation. Performance is assessed using metrics such as accuracy, AUROC, average precision and F1. In transductive settings, where drugs appearing in the test set are also present during training but in different pairings, performance across models approaches saturation. In this context, FG-DDI achieves leading results, though improvements over strong baselines are modest.
More pronounced differences are observed in inductive settings on DrugBank. Two partition schemes are considered, one in which both drugs in a test pair are unseen during training and another in which only one drug is unseen. In these scenarios, FG-DDI shows higher accuracy and F1 than competing approaches, while AUROC differences remain small. The results indicate improved balance between precision and recall when generalising to new compounds, even when ranking performance changes only slightly.
Functional Group Signals and Model Behaviour
Additional analyses examine how functional-group enrichment contributes to predictive behaviour. Enrichment scores highlight specific group combinations with elevated interaction propensity. Among the highest reported values are combinations involving heterocyclic motifs such as triazole and pyrrole, as well as imidazole or urea paired with pyrrole. Other notable pairs include tertiary alcohol with pyrrole, pyrimidine with pyrrole and tertiary amine with pyrrolidine. These patterns reflect recurrent substructure combinations within the interaction data.
A case analysis of a specific interaction type characterised by increased absorption and elevated serum concentration illustrates how enrichment-informed gates concentrate on combinations of amines and aromatic systems in one drug with ester, amide, ether or aromatic motifs in the other. The distribution of gate values suggests that the model leverages these chemically meaningful associations when forming predictions.
An ablation study further evaluates the contribution of intra- and inter-molecular enrichment components. In transductive evaluation, using either component alone performs slightly below the backbone model without functional-group enhancement, while combining both yields the strongest overall results. In inductive settings, the full configuration also attains the highest accuracy and F1 among its variants, indicating complementary effects between intra- and inter-molecular signals.
FG-DDI introduces functional-group enrichment into a dual-view graph neural network for drug–drug interaction prediction. By integrating curated medicinal-chemistry priors into both intra- and inter-molecular reasoning, the approach maintains strong performance in transductive settings and demonstrates clearer advantages in inductive scenarios involving unseen drugs. Analyses of enrichment patterns, per-interaction-type behaviour and ablation experiments support the contribution of functional-group signals to improved generalisation and more stable precision–recall balance. Reported limitations include reliance on discrete functional-group detection, pre-computed enrichment statistics and restriction to structure-based modelling, which does not account for factors beyond molecular graphs.
Source: Journal of Biomedical Informatics
Image Credit: iStock
