BOND is a context-aware benchmark for biomedical entity normalization, designed to evaluate how AI models interpret ambiguous biological terms (e.g., "macrophage", "RB", "Oligs") when provided with rich metadata context.

Unlike traditional string-matching benchmarks, BOND requires reasoning over biological context—including tissue, disease, organism, and development stage information—to accurately map author-provided annotations to standardized ontology labels.

Key Features

Context-Aware: Each normalization pair includes tissue, disease, organism, and development stage metadata
Multi-Entity: Covers cell types, tissues, diseases, assays, sex, development stage, and ethnicity
Cross-Modal: Includes both single-cell RNA-seq and spatial transcriptomics datasets
High Quality: Semantically validated with embedding similarity filtering and expert review

Data Distribution

Figure 1: BOND Benchmark Data Distribution. Panel A shows entity type distribution across 25,416 pairs from 85 datasets. Panel B displays dataset composition by organism and modality (72 single-cell, 13 spatial). Panel C shows tissue diversity (186 unique tissues). Panel D lists top contributing studies.

Dataset Statistics

Metric	Value
Total Pairs	25,416
Datasets	85
Tissues	186 unique types
Organisms	Homo sapiens, Mus musculus
Modalities	Single-Cell (72), Spatial (13)

Data Splits

Stratified 80/10/10 split supporting both zero-shot evaluation and supervised fine-tuning:

Split	Count	Percentage
Train	20,332	80%
Validation	2,542	10%
Test	2,542	10%

Data Fields

Field	Description
`author_term`	Raw annotation from source data (e.g., "RB", "Oligs", "10x")
`ontology_label`	Ground-truth standardized label (e.g., "rod bipolar cell")
`ontology_id`	Unique ontology identifier (e.g., `CL:0000782`)
`field_type`	Entity category: `cell_type`, `tissue`, `disease`, `assay`, `sex`, `development_stage`, `self_reported_ethnicity`
`tissue`	Biological tissue context (e.g., "retina", "lung")
`organism`	Source organism (Homo sapiens or Mus musculus)
`disease`	Disease state (e.g., "small cell lung carcinoma", "normal")
`assay`	Sequencing technology (e.g., "10x 3' v3", "Smart-seq2")
`dataset_id`	Source dataset identifier from CELLxGENE

Dataset Provenance

BOND aggregates data from 85 high-quality studies via CELLxGENE Census:

Modality	Human	Mouse	Total
Single-Cell	61	13	74
Spatial	9	2	11
Total	70	15	85

Representative Studies

Study	Pairs
Tabula Sapiens	6,842
Human Lung Cell Atlas	3,114
Human Retina Atlas	1,830
Mouse Embryonic Dev Atlas	1,663
Human Fetal Gene Atlas	1,326
MSK Ovarian Cancer Atlas	771
HCA Integration Atlas	681
Human Embryonic Limb Atlas	622

Creation Pipeline

Figure 2: Benchmark Creation Pipeline. Context-aware tuples extracted from 99 CELLxGENE datasets undergo embedding-based filtering (sim ≥ 0.40), tiered verification (auto-accept for high confidence, LLM judge for edge cases), and human review. Final benchmark contains 25,416 verified pairs across 85 datasets.

Usage

from datasets import load_dataset

dataset = load_dataset("pankajrajdeo/bond-benchmark-v1")

# Access splits
train = dataset["train"]
val = dataset["validation"]
test = dataset["test"]

# Example
sample = train[0]
print(f"Input: {sample['author_term']}")
print(f"Context: {sample['tissue']}, {sample['disease']}")
print(f"Target: {sample['ontology_label']} ({sample['ontology_id']})")

Citation

@dataset{bond_benchmark_2026,
  author = {Rajdeo, Pankaj},
  title = {BOND: Biomedical Ontology Normalization and Disambiguation},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/pankajrajdeo/bond-benchmark-v1}
}

License

This dataset is released under CC-BY-4.0.

Downloads last month: 34