SourceData Dataset

The largest annotated biomedical corpus for machine learning and AI in the publishing context.

SourceData is the largest annotated biomedical dataset for NER and NEL. It is unique on its focus on the core of scientific evidence: figure captions. It is also unique on its real-world configuration, since it does not present isolated sentences out of more general context. It offers full annotated figure captions that can be further enriched in context using full text, abstracts, or titles. The goal is to extract the nature of the experiments on them described. SourceData presents also its uniqueness by labelling the causal relationship between biological entities present in experiments, assigning experimental roles to each biomedical entity present in the corpus.

SourceData consistently annotates nine different biological entities (genes, proteins, cells, tissues, subcellular components, species, small molecules, and diseases). It is the first dataset annotating experimental assays and the roles played on them by the biological entities. Each entity is linked to their correspondent ontology, allowing for entity disambiguation and NEL.

Cite our work

@ARTICLE{2023arXiv231020440A,
       author = {{Abreu-Vicente}, Jorge and {Sonntag}, Hannah and {Eidens}, Thomas and {Lemberger}, Thomas},
        title = "{The SourceData-NLP dataset: integrating curation into scientific publishing for training large language models}",
      journal = {arXiv e-prints},
     keywords = {Computer Science - Computation and Language},
         year = 2023,
        month = oct,
          eid = {arXiv:2310.20440},
        pages = {arXiv:2310.20440},
archivePrefix = {arXiv},
       eprint = {2310.20440},
 primaryClass = {cs.CL},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2023arXiv231020440A},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}
@article {Liechti2017,
     author = {Liechti, Robin and George, Nancy and Götz, Lou and El-Gebali, Sara and Chasapi, Anastasia and Crespo, Isaac and Xenarios, Ioannis and Lemberger, Thomas},
     title = {SourceData - a semantic platform for curating and searching figures},
     year = {2017},
     volume = {14},
     number = {11},
     doi = {10.1038/nmeth.4471},
     URL = {https://doi.org/10.1038/nmeth.4471},
     eprint = {https://www.biorxiv.org/content/early/2016/06/20/058529.full.pdf},
     journal = {Nature Methods}
}

Dataset usage

The dataset has a semantic versioning. Specifying the version at loaded will give different versions. Below we is shown the code needed to load the latest available version of the dataset. Check below at Changelog to see the changes in the different versions.

  from datasets import load_dataset
  # Load NER
  ds = load_dataset("EMBO/SourceData", "NER", version="2.0.3")
  # Load PANELIZATION
  ds = load_dataset("EMBO/SourceData", "PANELIZATION", version="2.0.3")
  # Load GENEPROD ROLES
  ds = load_dataset("EMBO/SourceData", "ROLES_GP", version="2.0.3")
  # Load SMALL MOLECULE ROLES
  ds = load_dataset("EMBO/SourceData", "ROLES_SM", version="2.0.3")
  # Load MULTI ROLES
  ds = load_dataset("EMBO/SourceData", "ROLES_MULTI", version="2.0.3")

Note that we offer the XML serialized dataset. This includes all the data needed to perform NEL in SourceData. For reproducibility, for each big version of the dataset we provide split_vx.y.z.json files to generate the train, validation, test splits.

Supported Tasks and Leaderboards

Tags are provided as IOB2-style tags. PANELIZATION: figure captions (or figure legends) are usually composed of segments that each refer to one of several 'panels' of the full figure. Panels tend to represent results obtained with a coherent method and depicts data points that can be meaningfully compared to each other. PANELIZATION provide the start (B-PANEL_START) of these segments and allow to train for recogntion of the boundary between consecutive panel lengends. NER: biological and chemical entities are labeled. Specifically the following entities are tagged:

• SMALL_MOLECULE: small molecules
• GENEPROD: gene products (genes and proteins)
• SUBCELLULAR: subcellular components
• CELL_LINE: cell lines
• CELL_TYPE: cell types
• TISSUE: tissues and organs
• ORGANISM: species
• DISEASE: diseases (see limitations)
• EXP_ASSAY: experimental assays ROLES: the role of entities with regard to the causal hypotheses tested in the reported results. The tags are:
• CONTROLLED_VAR: entities that are associated with experimental variables and that subjected to controlled and targeted perturbations.
• MEASURED_VAR: entities that are associated with the variables measured and the object of the measurements.

In the case of experimental roles, it is generated separatedly for GENEPROD and SMALL_MOL and there is also the ROLES_MULTI that takes both at the same time.

Languages

The text in the dataset is English.

Dataset Structure

Data Instances

Data Fields

• words: list of strings text tokenized into words.
• panel_id: ID of the panel to which the example belongs to in the SourceData database.
• label_ids:
  • entity_types: list of strings for the IOB2 tags for entity type; possible value in ["O", "I-SMALL_MOLECULE", "B-SMALL_MOLECULE", "I-GENEPROD", "B-GENEPROD", "I-SUBCELLULAR", "B-SUBCELLULAR", "I-CELL_LINE", "B-CELL_LINE", "I-CELL_TYPE", "B-CELL_TYPE", "I-TISSUE", "B-TISSUE", "I-ORGANISM", "B-ORGANISM", "I-EXP_ASSAY", "B-EXP_ASSAY"]
  • roles: list of strings for the IOB2 tags for experimental roles; values in ["O", "I-CONTROLLED_VAR", "B-CONTROLLED_VAR", "I-MEASURED_VAR", "B-MEASURED_VAR"]
  • panel_start: list of strings for IOB2 tags ["O", "B-PANEL_START"]
  • multi roles: There are two different label sets. labels is like in roles. is_category tags GENEPROD and SMALL_MOLECULE.

Data Splits

• NER and ROLES

  DatasetDict({
      train: Dataset({
          features: ['words', 'labels', 'tag_mask', 'text'],
          num_rows: 55250
      })
      test: Dataset({
          features: ['words', 'labels', 'tag_mask', 'text'],
          num_rows: 6844
      })
      validation: Dataset({
          features: ['words', 'labels', 'tag_mask', 'text'],
          num_rows: 7951
      })
  })

• PANELIZATION

  DatasetDict({
      train: Dataset({
          features: ['words', 'labels', 'tag_mask'],
          num_rows: 14655
      })
      test: Dataset({
          features: ['words', 'labels', 'tag_mask'],
          num_rows: 1871
      })
      validation: Dataset({
          features: ['words', 'labels', 'tag_mask'],
          num_rows: 2088
      })
  })

Information Extraction

This folder contains caption-answer pairs intended to be used for information extraction. Each of the files contains answers to given questions about the captions. Each file is provided in csv and json format for convinience for different cases. The list of files and questions they answer are shown below:

• assayed_entities: What is the assayed/measured entity?

• chemicals: Are there any chemical compounds or small molecules mentioned?

• diseases: Is there any disease term mentioned, or can be infered, in the figure legend?

• experimental_assay: What kind of experimental assay was used for this experiment?

• hypothesis_tested: Can you formulate the hypothesis that this experiment has tested.

• is_experiment: Does the legend describe an experiment or not?

• ncbi_gene_linking: Can you link the identified genes to their NCBI gene identifiers?

• where_was_tested: In what kind of cell/tissue/organism/subcellular component was the experiment performed?

We refer the interested reader to the cypher queries used to generate this data for further information.

Compound Image Segmentation

This folder contain the data for the compound image segmentation task. The data is provided in format compatible to train YOLOv10 models.

The file soda_panelization_figures.zip contains 13039 figures extracted from scientific manuscripts, that are labeled to use object detection algorithms to separate the figure into its panels. The dataset is divided into train, validation and test sets.

The file segmented_images.zip contains panel-caption pairs. These have been used, together with multimodal LLMs to assign the correct panel label and caption to each panel in the figure.

Dataset Creation

Curation Rationale

The dataset was built to train models for the automatic extraction of a knowledge graph based from the scientific literature. The dataset can be used to train models for text segmentation, named entity recognition and semantic role labeling.

Source Data

Initial Data Collection and Normalization

Figure legends were annotated according to the SourceData framework described in Liechti et al 2017 (Nature Methods, 2017, https://doi.org/10.1038/nmeth.4471). The curation tool at https://curation.sourcedata.io was used to segment figure legends into panel legends, tag enities, assign experiemental roles and normalize with standard identifiers (not available in this dataset). The source data was downloaded from the SourceData API (https://api.sourcedata.io) on 21 Jan 2021.

Who are the source language producers?

The examples are extracted from the figure legends from scientific papers in cell and molecular biology.

Annotations

Annotation process

The annotations were produced manually with expert curators from the SourceData project (https://sourcedata.embo.org)

Who are the annotators?

Curators of the SourceData project.

Personal and Sensitive Information

None known.

Considerations for Using the Data

Social Impact of Dataset

Not applicable.

Discussion of Biases

The examples are heavily biased towards cell and molecular biology and are enriched in examples from papers published in EMBO Press journals (https://embopress.org)

The annotation of diseases has been added recently to the dataset. Although they appear, the number is very low and they are not consistently tagged through the entire dataset. We recommend to use the diseases by filtering the examples that contain them.

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

Thomas Lemberger, EMBO. Jorge Abreu Vicente, EMBO

Licensing Information

CC BY 4.0

Citation Information

We are currently working on a paper to present the dataset. It is expected to be ready by 2023 spring. In the meantime, the following paper should be cited.

  @article {Liechti2017,
      author = {Liechti, Robin and George, Nancy and Götz, Lou and El-Gebali, Sara and Chasapi, Anastasia and Crespo, Isaac and Xenarios, Ioannis and Lemberger, Thomas},
      title = {SourceData - a semantic platform for curating and searching figures},
      year = {2017},
    volume = {14},
    number = {11},
      doi = {10.1038/nmeth.4471},
      URL = {https://doi.org/10.1038/nmeth.4471},
      eprint = {https://www.biorxiv.org/content/early/2016/06/20/058529.full.pdf},
      journal = {Nature Methods}
  }

Contributions

Thanks to @tlemberger and @drAbreu for adding this dataset.

Changelog

• v2.0.3 - Data curated until 20.09.2023. Correction of 2,000+ unnormalized cell entities that have been now divided into cell line and cell type. Specially relevant for NER, not that important for NEL.

• v2.0.2 - Data curated until 20.09.2023. This version will also include the patch for milti-word generic terms.

• v1.0.2 - Modification of the generic patch in v1.0.1 to include generic terms of more than a word.

• v1.0.1 - Added a first patch of generic terms. Terms such as cells, fluorescence, or animals where originally tagged, but in this version they are removed.

• v1.0.0 - First publicly available version of the dataset. Data curated until March 2023.