Improving drug repositioning with negative data labeling using large language models.

Area of Science:

• Computational biology
• Drug discovery
• Machine learning

Background:

• Drug repositioning accelerates development and reduces costs but is limited by the scarcity of negative data (drugs failing due to inefficacy or toxicity).
• Existing Positive-Unlabeled (PU) learning methods struggle with misclassification and simplified decision boundaries due to challenges in accessing reliable negative drug data.

Purpose of the Study:

• To develop a novel strategy for systematically identifying true negative drugs using Large Language Models (LLMs).
• To enhance the accuracy and generalization of supervised machine learning models for drug repositioning.
• To identify potential drug candidates for prostate cancer treatment.

Main Methods:

• Utilized Large Language Models (GPT-4) to analyze clinical trials for prostate cancer and identify true negative drugs.
• Created a training dataset of 26 positive and 54 validated negative drugs.
• Applied a machine learning ensemble to screen 11,043 drugs from the DrugBank database for repurposing potential.

Main Results:

• The LLM-based strategy significantly improved predictive accuracy (Matthews Correlation Coefficient of 0.76) compared to traditional PU learning approaches (0.55 and 0.48).
• Identified 980 potential drug candidates for prostate cancer from 11,043 screened drugs.
• Detailed review of the top 30 candidates revealed 9 promising drugs targeting key mechanisms like genomic instability and p53 regulation.

Conclusions:

• The developed negative data labeling approach using LLMs can substantially advance supervised drug repositioning.
• Expanding this method to all diseases in ClinicalTrials.gov offers a more accurate, data-driven approach to discovering new therapeutics.
• This strategy promises to accelerate the identification of effective treatments by overcoming limitations in traditional drug development pipelines.