Models and Processes to Extract Drug-like Molecules From Natural Language Text

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Drug repurposing and discovery are critical for combating SARS-CoV-2.
• Scientific literature contains vast information on potential drug-like molecules.
• Existing Named Entity Recognition (NER) models struggle with specialized vocabularies in viral research literature.

Purpose of the Study:

• To develop and evaluate a human-AI hybrid pipeline for identifying drug-like molecules from scientific text.
• To address the limitations of existing NER models in processing large, specialized scientific corpora.
• To accelerate the identification of potential drug candidates for SARS-CoV-2 research.

Main Methods:

• Implemented an iterative model-in-the-loop strategy for efficient training data generation.
• Utilized scarce human expertise judiciously to train a specialized NER model.
• Applied the developed NER model to the COVID-19 Open Research Dataset Challenge (CORD-19) corpus.

Main Results:

• Achieved an F-1 score of 80.5% for the NER model, comparable to human performance.
• Identified 10,912 putative drug-like molecules within the CORD-19 corpus.
• Enriched computational screening targets by 3,591 molecules, with 18 ranking in the top 0.1% for 3CLPro docking.

Conclusions:

• The human-AI hybrid pipeline effectively identifies drug-like molecules from large scientific datasets.
• This approach significantly enhances the efficiency of drug discovery for viral diseases like COVID-19.
• The method offers a scalable solution for extracting valuable molecular information from scientific literature.