Semantic Priors for Drug–Drug Interaction Prediction Using Compact Graph Encoders

Accurate prediction of drug–drug interactions (DDI) is critical to patient safety. Graph-based models show promise for DDI link prediction, with prior work exploring both structure-only encoders and those augmented with semantic information. However, there is limited evaluation of whether semantic priors enable smaller encoders to reach the performance of larger models. We investigate whether domain-aware biomedical text embeddings, that are task-optimized and used for node initialization, enable compact encoders to achieve predictive accuracy comparable to that of much larger graph-based models. We precompute drug node embeddings by encoding DrugBank text with a frozen SciBERT, refine these embeddings with a small contrastive MLP, and use the resulting task-oriented embeddings to initialize node representations in a Graph Neural Network. During training, the model learns structural information on top of this semantic prior, with node embeddings regularized to remain close to their initial values. On the ogbl-ddi benchmark, our model attains test performance approaching the best published structure-only result, while using only 0.52\% of the parameters (5.08 million vs. 976 million). Among published models on ogbl-ddi, our approach lies on the Pareto frontier of performance versus size and outperforms 94\% of the existing entries. These findings suggest that semantic priors from pretrained scientific language models with task-optimized refinement can support resource-efficient, competitive encoders for DDI link prediction.