Scholarlink

Co-administration of multiple drugs is often recommended in clinical treatments to enhance therapeutic efficacy. However, inappropriate drug combinations can lead to adverse side effects and diminished treatment outcomes. Therefore, accurate prediction of drug-drug interactions (DDI) remains a critical challenge in pharmacology and patient safety. Numerous computational and machine learning approaches are available and treat DDI prediction as a binary classification task or predict specific interaction categories. However, recent advancements in large language models (LLMs) offer new opportunities, particularly in generating human-readable explanations in biomedical sciences. In this study, we propose a novel framework for generating descriptive, natural language explanations of DDIs using a sequence-to-sequence MOLT5 transformer model. Our approach leverages three variants of pre-trained MOLT5 models (i.e., small, base, and large), and uniquely integrates both the common names of drugs and their corresponding SMILES representations. This dual-modality input enables the model to capture both semantic and structural information of drug compounds. The models were trained on a curated dataset of drug pairs and their interaction descriptions from DrugBank. Among the tested variants, the large MOLT5 model achieved the best performance, with ROUGE-1: 82.92, ROUGE-2: 69.67, ROUGE-L: 78.34, and BLEU: 68.87 on the validation set. Additionally, we enhanced the model’s prediction by integrating structural features from drug molecular graphs, further improving the representation and interpretability of DDI predictions. Our proposed framework demonstrates the potential of LLMs to advance the field of pharmacovigilance by generating accurate and interpretable DDI descriptions with important structural features. The model is available at https://huggingface.co/acdsd/DDI .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating Large Language Models and Explainable AI for the Efficient Drug-Drug Interactions Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Gori Sankar Borah,
  • Amit Kalita,
  • Debasish Saikia,
  • Selvaraman Nagamani

摘要

Co-administration of multiple drugs is often recommended in clinical treatments to enhance therapeutic efficacy. However, inappropriate drug combinations can lead to adverse side effects and diminished treatment outcomes. Therefore, accurate prediction of drug-drug interactions (DDI) remains a critical challenge in pharmacology and patient safety. Numerous computational and machine learning approaches are available and treat DDI prediction as a binary classification task or predict specific interaction categories. However, recent advancements in large language models (LLMs) offer new opportunities, particularly in generating human-readable explanations in biomedical sciences. In this study, we propose a novel framework for generating descriptive, natural language explanations of DDIs using a sequence-to-sequence MOLT5 transformer model. Our approach leverages three variants of pre-trained MOLT5 models (i.e., small, base, and large), and uniquely integrates both the common names of drugs and their corresponding SMILES representations. This dual-modality input enables the model to capture both semantic and structural information of drug compounds. The models were trained on a curated dataset of drug pairs and their interaction descriptions from DrugBank. Among the tested variants, the large MOLT5 model achieved the best performance, with ROUGE-1: 82.92, ROUGE-2: 69.67, ROUGE-L: 78.34, and BLEU: 68.87 on the validation set. Additionally, we enhanced the model’s prediction by integrating structural features from drug molecular graphs, further improving the representation and interpretability of DDI predictions. Our proposed framework demonstrates the potential of LLMs to advance the field of pharmacovigilance by generating accurate and interpretable DDI descriptions with important structural features. The model is available at https://huggingface.co/acdsd/DDI .