Enhanced Drug-drug Interaction Prediction Using Adaptive Knowledge Integration

Abstract

Drug-drug interaction event (DDIE) prediction is crucial for preventing adverse reactions and ensuring optimal therapeutic outcomes. However, existing methods often face challenges with imbalanced datasets, complex interaction mechanisms, and poor generalization to unknown drug combinations. To address these challenges, we propose a knowledge augmentation framework that adaptively infuses prior drug knowledge into a large language model (LLM). This framework utilizes reinforcement learning techniques to facilitate adaptive knowledge extraction and synthesis, thereby efficiently optimizing the strategy space to enhance the accuracy of LLMs for DDIE predictions. As a result of few-shot learning, we achieved a notable improvement compared to the baseline. This approach establishes an effective framework for scientific knowledge learning for DDIE predictions.

Figures (6)

• Figure 1: Overview of the framework. Leveraging the information of drug molecules to enhance the accuracy of DDIE prediction in few-shot scenarios. (a) Knowledge Extraction classifies drug types from features using specific clustering methods and counts. (b) Prompt Manager uses drug types as prior knowledge and combines them with molecular representations or descriptions to form prompts. (c) LLM-based DDIE Predictor predicts DDIEs, with the performance feedback serving as a reward to the Adaptive Strategy Selector. (d) Adaptive Strategy Selector develops strategies based on clustering methods, knowledge synthesis strategies, and LLM hyperparameters, exploring the strategy space to iteratively select optimal states.
• Figure 2: Dataset analysis. (a) and (b) show the frequency distribution of DDIE categories in the DeepDDI 2 and DDIMDL datasets. Categories are grouped into Common, Few, and Rare based on sample frequency, with the y-axis showing the number of samples and the x-axis showing the category indices. Note that although some numerical labels may overlap across datasets, the underlying categories differ.
• Figure 3: Accuracy performance of baseline models and ours on DeepDDI 2 (a) and DDIMDL (b) datasets across data splits (All Events, Common, Few, Rare).
• Figure 4: The exploration process of strategy combinations for the few and rare data splits in DeepDDI 2 dataset. The horizontal axis represents the exploration steps, while the vertical axis shows the F1-score evaluation metric. Different colors indicate the corresponding clustering methods. The top-3 F1 scores in each split are highlighted with red circles. The right axes show the corresponding Reward (purple dashed line) and Validation Loss (crimson dash-dot line).
• Figure 5: Ablation study on DeepDDI 2 (a) and DDIMDL (b) datasets for Few and Rare splits. "GR" denotes coarse grid search, "RS" denotes random search, and "Ours" is the proposed RL-based searcher. Error bars indicate standard deviation over three runs.