An LLM-Driven Pipeline for Automated Quantitative Drug-Target Interaction Data Extraction and Curation from Biomedical Literature

The accurate understanding of quantitative drug-target interactions (DTIs) is crucial for modern drug discovery, yet a vast amount of essential affinity data (IC 50 , K i , K d , EC 50) remains embedded within unstructured biomedical literature, leading to sparse and biased datasets for machine learning models. To overcome this challenge, we introduce PharmExtract, an automated, endto-end pipeline leveraging large language models (LLMs), optical character recognition (OCR), and advanced table parsing to systematically extract, standardize, and curate high-quality quantitative DTI data. Processing hundreds of thousands of full-text articles, PharmExtract yields a comprehensive collection of unique DTI entries after rigorous quality control, notably including a significant proportion of novel data absent from existing databases. Our evaluation demonstrates PharmExtract achieves significantly higher extraction accuracy and precision compared to baseline methods. Furthermore, retraining established DTI prediction models (GraphDTA, DeepDTA, MolTrans) with PharmExtractDB consistently demonstrates substantial performance improvements. These results highlight PharmExtractDB's transformative potential in providing enriched, high-quality DTI data, thereby accelerating the development of more accurate DTI prediction models and advancing computational drug discovery.