Large-Scale Meta-Analysis of Nanomaterials Toxicity Based on Natural Language Processing of Scientific Articles

Abstract

Natural language processing (NLP) pipelines can mine the nanotoxicology literature at a scale and resolution that cannot be achieved by manual curation. Here, we established a NLP pipeline that coupled topic modeling and end point-specific extraction LLM prompts to convert ∼106 sentences extracted from abstracts of scientific articles into a structured knowledge base containing 13 nanotoxicology features. The pipeline is capable of analyzing and extracting nanomaterial descriptors such as size, ζ potential, and surface area, along with biological end points such as minimum inhibitory concentration (MIC), minimum bactericidal concentration (MBC) and lethal concentration 50% (LC50). Statistical convergence across multiple quantitative end points - MIC, MBC, microbial log reduction, and biofilm killing efficiency - shows that Ag-based nanomaterials are the most potent antimicrobial agents, showing lower MIC and MBC than ZnO, TiO2 and Au analogs. This trend was also observed for individual pathogens such as Escherichia coli and Staphylococcus aureus. Most nanomaterials are within 1 to 100 nm, with nanoparticles featured in >80% of the studies. Although nanomaterials <50 nm often produce the lowest MIC and LC50, toxicity within a single size class spans orders of magnitude, underscoring the influence of surface chemistry, coatings, and colloidal behavior. In addition, adverse reproductive effects in Caenorhabditis elegans and Daphnia magna, and developmental abnormalities in Danio rerio, are predominantly related to Ag and TiO2. In general, our automated data extraction and the curation strategy transforms disparate literature into a machine-readable knowledge base that paves the way for data-driven predictions of nanomaterial hazards.