The CellLink Corpus: A text dataset of cells in the biomedical literature

This Zenodo repository contains the data described in our manuscript, "Cell phenotypes in the biomedical literature: a systematic analysis and text mining corpus", doi:10.64898/2026.02.11.705457.

Data overview:
Single-cell technologies have recently enabled the discovery of many novel cell phenotypes. However, the complex assortment of cell phenotypes described in these studies remains fragmented across the scientific literature. The repository described here contains the CellLink corpus—a set of excerpts from recent human and mouse biomedical articles. The corpus was manually annotated by expert curators according to the annotation guidelines provided, to support the development and evaluation of machine learning models for automatically identifying cell types in the literature, including novel cell types. The excerpts (“passages”) included were selected from both abstracts and full-text articles published in 2019-2024 which focus on single-cell technologies or cell types.

The corpus contains 22,360 annotations across 3,005 passages extracted from 2,765 articles in 467 journals. The corpus distinguishes 3 types of mentions: 83% of the annotations are cell phenotypes, 7% are heterogeneous cell populations, and 10% are vague cell populations. Mentions of the first two types are linked to Cell Ontology (v2025-01-08) identifiers where possible. Linking is with respect to the meaning of the mention in context, with matches labeled as exact or related. The source articles for all excerpts from full-text articles are licensed under CC-BY or CC0; CellLink annotations are released into the public domain.

The repository contains the selected passages and manual annotations for the training and validation sets, provided in BioC XML format. The repository also contains the unannotated articles for the held-out test set; the annotations will be available (indirectly) to the public via CodaBench, to support fair evaluation. Users can upload predictions for the test set and receive an evaluation, as described on CodaBench. The evaluation script is included in this repository. A detailed description of the corpus is available in our manuscript, "Cell phenotypes in the biomedical literature: a systematic analysis and text mining corpus". This record will be updated as the additional releases become public.

Acknowledgment:
This research was supported by the Intramural Research Program of the National Institutes of Health (NIH). The contributions of the NIH authors are considered Works of the United States Government. The findings and conclusions presented in this paper are those of the authors and do not necessarily reflect the views of the NIH or the U.S. Department of Health and Human Services.

Example code minisnippet:

"""
Print the first annotation in the CellLink train set.
"""

import bioc

filepath = "train.xml" # edit here

# load the bioc-xml file
with open(filepath, 'r', encoding='utf-8') as readfp:
    bioc_collection = bioc.load(readfp)

first_document = bioc_collection.documents[0]
first_passage = first_document.passages[0]
first_annotation = first_passage.annotations[0]
print(first_annotation)

>> BioCAnnotation[id=0,text='progenitor cells',infons=[type=cell_phenotype,identifier=(skos:exact)CL:0011026],locations=[BioCLocation[offset=317,length=16]],]

Additional example programs are provided.