Improving drug-drug interaction prediction via in-context learning and judging with large language models.

Introduction: Large Language Models (LLMs), recognized for their advanced capabilities in natural language processing, have been successfully employed across various domains. However, their effectiveness in addressing challenges related to drug discovery has yet to be fully elucidated. Methods: In this paper, we propose a novel LLM based method for drug-drug interaction (DDI) prediction, named DDI-JUDGE, achieved through the integration of judging and ICL prompts. The proposed method outperforms existing LLM approaches, demonstrating the potential of LLMs for predicting DDIs. We introduce a novel in-context learning (ICL) prompt paradigm that selects high-similarity samples as positive and negative prompts, enabling the model to effectively learn and generalize knowledge. Additionally, we present an ICL-based prompt template that structures inputs, prediction tasks, relevant factors, and examples, leveraging the pre-trained knowledge and contextual understanding of LLMs to enhance DDI prediction capabilities. To further refine predictions, we employ GPT-4 as a discriminator to assess the relevance of predictions generated by multiple LLMs. Results: DDI-JUDGE achieves the best performance among all models in both zero-shot and few-shot settings, with an AUC of 0.642/0.788 and AUPR of 0.629/0.801, respectively. These results demonstrate its superior predictive capability and robustness across different learning scenarios. Development: These findings highlight the potential of LLMs in advancing drug discovery through more effective DDI prediction. The modular prompt structure, combined with ensemble reasoning, offers a scalable framework for knowledge-intensive biomedical applications. The code for DDI-JUDGE is available at https://github.com/zcc1203/ddi-judge.