Large Language Model-Based Natural Language Encoding Could Be All You Need for Drug Biomedical Association Prediction

Area of Science:

• Biomedicine
• Artificial Intelligence
• Drug Discovery

Background:

• Analyzing drug-related interactions is crucial for drug discovery and development.
• Existing AI tools for drug biomedical associations (DBAs) lack comprehensive feature encoding for biomedical functions and semantic concepts.
• Large language models (LLMs) show promise due to their advanced natural language understanding.

Purpose of the Study:

• To introduce LEDAP, a novel method leveraging LLM-based biotext feature encoding for predicting drug associations.
• To evaluate LEDAP's performance in analyzing drug-disease associations, drug-drug interactions, and drug-side effect associations.
• To demonstrate the potential of LLMs in advancing drug development analysis.

Main Methods:

• Developed LEDAP, a system utilizing LLM-based feature encoding for DBA prediction.
• Employed LLMs for their holistic understanding of natural language and biomedical topics.
• Integrated LLM-based feature representations with classical machine learning methods.

Main Results:

• LEDAP demonstrated competitive performance compared to existing DBA analysis tools.
• LLM-based feature representations achieved satisfactory performance across various DBA tasks, including binary classification, multiclass classification, and regression.
• The approach showed consistent effectiveness even with simple machine learning models.

Conclusions:

• LLMs possess considerable potential for drug development research.
• LEDAP's approach offers a significant advancement in analyzing drug biomedical associations.
• The findings suggest LLMs can act as a catalyst for future progress in drug discovery and development.