ant-p/cancer-mirna-18

Extracellular miRNA Expression for Cancer Diagnosis

This dataset contains extracellular microRNA (miRNA) expression profiles for 25,627 human samples covering 17 tumor types and one healthy control class. It integrates 53 microarray studies into a single, harmonized feature space.

It was built for the paper "Boosting Heterogeneous Graph Attention Networks for Cancer Diagnosis from Extracellular microRNA Expression Data" (under review at Bioinformatics), where it is used to evaluate the HBoost-GAT method.

Dataset Summary

Extracellular miRNAs circulate in biofluids such as blood and serum. They are stable and carry disease-specific expression patterns, which makes them promising non-invasive biomarkers for cancer diagnosis through liquid biopsy.

Public miRNA data are spread across many studies that use different platforms, identifiers, normalizations, and label conventions, so they cannot be combined directly. This dataset provides a pre-processed and harmonized collection that can be used directly for machine learning.

Files

File	Content
EXP_microArraySeq.csv	Expression matrix. One row per sample, with a sample_id column followed by one column per miRNA.
PD_microArraySeq.csv	Phenotype data (sample metadata). One row per sample, including the disease label.

The two files can be joined on sample_id.

Expression file (EXP_microArraySeq.csv)

• sample_id: sample identifier.
• One column per miRNA (2,735 features). Names follow miRBase v22 nomenclature, for example hsa-miR-21-5p or hsa-let-7a-5p. Most columns are mature miRNAs. Some columns use precursor identifiers (lowercase hsa-mir-..., or hsa-let-7* without an arm suffix), as reported by the original studies.
• Values are log2(x + 1) normalized expression levels and are non-negative.

Missing values have a specific meaning:

Value	Meaning
0	The miRNA was on the study's array panel, but no expression value was reported. It is treated as not expressed.
empty	The miRNA was not on the study's panel, so it was never measured for that sample.

Because each study used its own panel, empty cells form study-specific blocks: all samples from the same study share the same set of measured miRNAs. Treat empty cells as truly missing, not as zero expression. In the paper, empty values were replaced with 0 only when applying PCA.

Metadata file (PD_microArraySeq.csv)

Column	Description
sample_id	GEO sample accession (GSM...). It matches sample_id in the expression file and can be used to trace a sample back to its original study on GEO.
technology	Microarray platform type: spotted oligonucleotide (23,329 samples) or in situ oligonucleotide (2,298).
sample_type	Origin of the profiled RNA: circulating (24,142), exosomes EV (1,395), not-exosomes EV (68), or whole cell (22).
type	Source material: serum (24,250), biofluids (722), plasma (385), whole blood (163), or cell (85). It is empty for the 22 whole cell samples.
gender	M (8,487), F (6,181), or uncertain (240). It is empty when the original study did not report it (10,719).
age	Age in years (range 1–100, median 62). It is empty when not reported (9,850).
disease	Standardized disease label. This is the target variable.

Class Distribution

Disease	Samples
Healthy control	14,122
Lung cancer	2,184
Breast cancer	2,135
Prostate cancer	1,407
Sarcoma	1,200
Esophageal cancer	826
Hepatocellular carcinoma	692
Bladder cancer	442
Ovarian cancer	433
Pancreatic cancer	429
Colon/Colorectal cancer	413
Gastric cancer	358
Glioma	307
Biliary tract cancer	254
Liver cancer	163
Nasopharyngeal carcinoma	132
Borderline ovarian tumor	66
Wilms tumor	64
Total	25,627

The dataset is strongly imbalanced. Healthy controls make up about 55% of the samples, and only 6 of the tumor classes have more than 500 samples. Macro-averaged metrics are recommended for evaluation. The table uses the exact label strings found in the disease column.

Data Collection and Processing

The raw data come from ExomiRHub, a platform that integrates extracellular miRNA studies deposited in NCBI GEO. The starting collection had 191 studies and 29,198 samples. Processing had four phases:

1. Study filtering. Studies unrelated to cancer were removed. Redundant studies, meaning subsets of other studies, were also removed. Only microarray-based studies were kept, and NGS and less standardized platforms were excluded.
2. Data alignment. miRNA identifiers were converted to canonical miRBase v22 names with miRBaseConverter. Identifiers that could not be mapped were discarded. Disease labels were standardized through manual inspection of study metadata.
3. Sample filtering. The following samples were removed: non-human samples, samples collected with polyA RNA selection, samples from two-channel microarrays, and samples with an undeterminable or non-oncological disease condition.
4. Merging. Studies were merged into a single feature space made of all miRNAs measured in at least one study. Studies normalized with log2(x) were re-normalized to log2(x + 1), which gives a consistent, non-negative range. Missing values were handled as described above. Classes with fewer than 50 samples were discarded.

Usage

from datasets import load_dataset

meta = load_dataset("ant-p/cancer-mirna-18", "metadata", split="full").to_pandas()
expr = load_dataset("ant-p/cancer-mirna-18", "expression", split="full").to_pandas()

df = meta.merge(expr, on="sample_id")
X = df[expr.columns.drop("sample_id")]
y = df["disease"]

Intended Uses and Limitations

• Intended use: research on machine learning methods for cancer classification and biomarker discovery from extracellular miRNA data.
• Not for clinical use. The dataset has not been validated for diagnostic decisions on individual patients.
• Privacy: all data derive from publicly available, de-identified GEO submissions.

License

This dataset is released under CC BY 4.0. It is derived from public GEO studies collected through ExomiRHub. Users should also acknowledge the original data sources.

Citation

If you use this dataset, please cite:

@unpublished{pellicani2026hboostgat,
  title  = {Boosting Heterogeneous Graph Attention Networks for Cancer Diagnosis from Extracellular microRNA Expression Data},
  author = {Pellicani, Antonio and Detomaso, Gabriele and Tavoletti, Elena and Pio, Gianvito and Ceci, Michelangelo},
  note   = {Under review at Bioinformatics},
  year   = {2026}
}

Please also cite ExomiRHub:

@article{liu2024exomirhub,
  title   = {ExomiRHub: A comprehensive database for hosting and analyzing human disease-related extracellular microRNA transcriptomics data},
  author  = {Liu, Yang and Min, Zhuochao and Mo, Jing and Ju, Zhen and Chen, Jianliang and Liang, Weiling and Zhang, Lantian and Li, Hanguang and Chan, Godfrey Chi-Fung and Wei, Yanjie and Zhang, Wenliang},
  journal = {Computational and Structural Biotechnology Journal},
  volume  = {23},
  pages   = {3104--3116},
  year    = {2024},
  doi     = {10.1016/j.csbj.2024.07.024}
}

Contact

Antonio Pellicani, antonio.pellicani@uniba.it. Department of Computer Science, University of Bari Aldo Moro.