CellExLink: End-to-end cell-type recognition and normalization in biomedical text

Cell types are described in biomedical literature using diverse names, abbreviations, and phenotype phrases, which complicates their recognition and normalization. We developed CellExLink, an end-to-end pipeline that identifies cell-type mentions and normalizes them to Cell Ontology (CL) identifiers. The recognizer was fine-tuned and evaluated on five heterogeneous biomedical corpora spanning full-length articles, article excerpts, figure captions, abstracts, and anatomical text passages. These resources include fine-grained phenotype-defined populations, heterogeneous cell populations, and abbreviated mentions. Across the five corpora, CellExLink achieved macro-average exact- and relaxed-span F1 scores of 0.766 and 0.855, respectively. For CL identifier normalization on gold-standard mention spans, F1 scores ranged from 0.690 to 0.874. In strict end-to-end evaluation, which required both an exact mention span and the correct CL identifier, F1 scores ranged from 0.552 on a figure-caption corpus to 0.725 on a corpus of full-text article excerpts. CellExLink outperformed the evaluated off-the-shelf systems in cell mention recognition, CL identifier normalization, and end-to-end extraction. By converting unannotated biomedical text into cell-type spans linked to standardized CL identifiers, CellExLink provides a practical foundation for downstream applications, including literature curation, relation extraction, and knowledge graph construction.