Integrating artificial intelligence and manual curation to enhance bioassay annotations in ChEMBL

Abstract As the volume and diversity of bioactivity data in ChEMBL continues to grow, ensuring that assay metadata is standardized, interoperable, and machine-readable is critical for effective use in cheminformatics and ML applications. In this work, we present recent efforts to enhance the quality and granularity of bioassay annotations in ChEMBL through a combination of manual and semi-manual curation and AI-driven approaches. We introduce a “perfect assay description” template to guide consistent annotation and demonstrate how natural language processing techniques and multi-class classification can be used to automatically extract key assay parameters and assign broad assay categories for legacy data. We report on the development, validation, and application of a spaCy-based NER model that identifies experimental methods with high precision and recall, as well as a complementary classification model that refines ASSAY_TYPE categorization beyond the existing schema. In addition, we describe improvements to metadata extraction for ADME endpoints, organism and protein variant annotations, and ontology linking using tools such as text2term. Together, these enhancements significantly advance the FAIRness of ChEMBL’s bioassay data, enabling more robust downstream analyses and more precise compound-target activity modeling. Graphical Abstract