databio/cell-line-nli

Dataset Card for BioTermNLI

This dataset is created with a structure similar to AllNLI. It is intended for fine-tuning embedding models to generate meaningful embeddings of biomedical terms, primarily consisting of alphanumeric identifiers such as K562, HepG2, and ZNF148.

Raw Data

From cellosaurus, we collected metadata of 30,001 cell lines that:

• Are human cell lines (NCBI_TaxID=9606)
• Have reference in BTO, CLO, or EFO
• With at least 1 of those annotations available:
  • Diseases: Prostate Carcinoma, Glioblastoma, etc.
  • Derived from site: lung primordium, corneal epithelium, etc.
  • Transformant: N-methyl-N-nitrosourea, Epstein Barr Virus, etc.
  • Cell type: kidney epithelial cell, skin fibroblast, etc.

triplet-terms subset

For each annotation type,

pair-class subset

Premise

each cell line and each of its available annotation (4 attributes plus category)

        if annotation_type == "category":
        # Paraphrases for cell line categories (e.g., "Cancer cell line", "Hybrid cell line")
        return random.choice(
            [
                f"{cell_line_term} is a type of {annotation_term}.",
                f"{cell_line_term} cell line belongs to the category of {annotation_term}.",
                f"The {annotation_term} includes {cell_line_term}.",
            ]
        )
    elif annotation_type == "cell_type":
        # Paraphrases for cell types (e.g., "aortic endothelial cell", "retinoblast")
        return random.choice(
            [
                f"{cell_line_term} is {annotation_term}.",
                f"{annotation_term} is the cell type for {cell_line_term}.",
                f"{cell_line_term} represents a {annotation_term}.",
            ]
        )
    elif annotation_type == "disease":
        # Paraphrases for diseases (e.g., "Cutaneous Melanoma", "Chronic granulomatous disease")
        return random.choice(
            [
                f"{cell_line_term} comes from a patient with {annotation_term}.",
                f"A patient of {annotation_term} contributed to {cell_line_term}.",
                f"The donor of {cell_line_term} has {annotation_term}.",
            ]
        )
    elif annotation_type == "transformant":
        # Paraphrases for transformants (e.g., "Ad12-SV40 hybrid virus", "Human papillomavirus 38")
        return random.choice(
            [
                f"The {annotation_term} transformed {cell_line_term}.",
                f"{annotation_term} is the transformant in the {cell_line_term} cell line.",
                f"The {cell_line_term} cell line contains {annotation_term} as a transformant.",
            ]
        )
    elif annotation_type == "site":
        # Paraphrases for sites (e.g., "bone marrow", "spleen")
        return random.choice(
            [
                f"{cell_line_term} was derived from {annotation_term}.",
                f"{annotation_term} is the origin of the {cell_line_term}.",
                f"{cell_line_term} is sourced from {annotation_term}.",
            ]
        )

Hypotheses

Entailment

Same cell line and annotation type, replaced them with synonyms if available (from cellosaurus and the ontology where the annotation comes from). Paraphrase by rule-based method.

Neutral

Same cell line, different annotation type

Contradiction

Different cell line, same annotation type

or:

Same cell line, same annotation type, wrong term?