electricsheepafrica/tumor-gene-expression-african

Tumor Gene Expression Panels in African Breast Cancer

Dataset owner: Electric Sheep Africa
Dataset type: Synthetic tumor gene expression (Oncotype DX, PAM50, Ki-67)
Populations: African and African-descent breast cancer patients
Version: 1.0.0
License: CC-BY-NC-4.0

---

1. Dataset Description

This dataset provides synthetic tumor gene expression profiles for 50,000 breast cancer cases from African and African-descent populations, designed to mirror:

• Oncotype DX–equivalent recurrence scores for ER+/HER2− tumors
• PAM50 molecular subtypes (Luminal A/B, Basal-like, HER2-enriched, Normal-like)
• Proliferation markers, including Ki-67 and related genes

All records are fully synthetic and were generated using an internal, literature-driven synthetic data methodology, with parameters derived from peer-reviewed literature on African and African-descent breast cancer cohorts.

Important: This dataset contains no real patient data. It is derived entirely from literature-based distributions and coherence rules.

---

2. Intended Use

This dataset is intended for:

• Method development for molecular subtyping (PAM50, IHC surrogates)
• Algorithm training for Oncotype DX–like recurrence score prediction
• Proliferation and Ki-67 modeling across African populations
• Health equity research on differences in tumor biology by ancestry
• Education and benchmarking in computational oncology

Not intended for clinical decision-making or individual risk prediction.

---

3. Populations & Cohort Design

3.1 Populations

The dataset includes 5 broad population groups, aligned with prior projects in this series:

• West_Africa (e.g., Nigeria, Ghana, Senegal)
• East_Africa (e.g., Kenya, Uganda, Ethiopia)
• Southern_Africa (e.g., South Africa, Namibia, Botswana)
• Central_Africa (e.g., Cameroon, DRC)
• African_American (USA)

3.2 Sample Size

• Total samples: 50,000 synthetic tumors
• Approximate population fractions:
  • West_Africa: 25%
  • East_Africa: 20%
  • Southern_Africa: 15%
  • Central_Africa: 10%
  • African_American: 30%

---

4. Molecular Features

4.1 PAM50 Molecular Subtypes

Each sample is assigned a PAM50-like molecular subtype:

• Luminal_A
• Luminal_B
• Basal_like
• HER2_enriched
• Normal_like

The distribution is African-enriched for basal-like tumors, reflecting the literature:

• Basal_like: ~40%
• Luminal_A: ~35%
• Luminal_B: ~17%
• HER2_enriched: ~8%
• Normal_like: ~1–2%

4.2 Receptor Status

Immunohistochemistry-style receptor status is included:

• ER_status: Positive / Negative
• PR_status: Positive / Negative
• HER2_status: Positive / Negative
• is_TNBC: Triple-negative (ER−/PR−/HER2−)

Triple-negative tumors (TNBC) are enriched (~35–40%) to reflect African cohorts.

4.3 Gene Expression Features (Key Subset)

All expression values are on a log2(normalized expression + 1) scale.

This initial public version exposes a key subset of genes for transparency and compactness:

• Proliferation marker:
  • MKI67_expr – Ki-67
• Hormone receptor signaling:
  • ESR1_expr – Estrogen receptor
  • PGR_expr – Progesterone receptor
• HER2 pathway:
  • ERBB2_expr – HER2 receptor
• Basal markers:
  • KRT5_expr, KRT17_expr – Basal cytokeratins

Future versions may expose additional genes from the Oncotype DX and PAM50 panels in a companion dataset.

4.4 Proliferation (Ki-67)

• ki67_percentage – Estimated Ki-67 labeling index (0–100%)
• ki67_category – {Low (<14%), Intermediate (14–30%), High (>30%)}

African and African-descent populations show higher Ki-67, particularly in basal-like and high-grade tumors.

4.5 Oncotype DX–Equivalent Score

For ER+/HER2− tumors, we provide a simplified Oncotype DX–like recurrence score:

• oncotype_RS – Integer score from 0–100
• oncotype_risk_category – {Low, Intermediate, High, Not_applicable}

The risk distribution is calibrated from published literature on African American and multi-ethnic cohorts.

---

5. Main File Schema

5.1 gene_expression_data.csv

• Rows: 50,000 samples
• Columns (23 variables):

Demographics & clinical:

• sample_id – Synthetic ID (BC_EXPR_00000 ...)
• population – One of 5 population groups
• age – Age in years (25–85)
• age_group – {<40, 40-49, 50-59, 60+}
• menopausal_status – {Premenopausal, Postmenopausal}
• bmi – Body mass index (kg/m²)
• tumor_grade – {1, 2, 3}
• stage – {I, II, III}

Molecular subtypes and receptors:

• pam50_subtype – {Luminal_A, Luminal_B, Basal_like, HER2_enriched, Normal_like}
• ER_status – {Positive, Negative}
• PR_status – {Positive, Negative}
• HER2_status – {Positive, Negative}
• is_TNBC – Boolean (triple-negative)

Key gene expression values (log2 scale):

• MKI67_expr – Ki-67
• ESR1_expr – Estrogen receptor
• PGR_expr – Progesterone receptor
• ERBB2_expr – HER2 receptor
• KRT5_expr, KRT17_expr – Basal cytokeratins

Derived scores:

• ki67_percentage – Ki-67 proliferation index (%)
• ki67_category – {Low, Intermediate, High}
• oncotype_RS – Oncotype-like recurrence score (0–100, NaN if not applicable)
• oncotype_risk_category – {Low, Intermediate, High, Not_applicable}

---

6. Data Access & Files

Main Dataset (root)

• gene_expression_data.csv – 50,000 × 23 variables (main table, CSV)
• gene_expression_data.parquet – Same table in Parquet format for efficient loading.

Auxiliary Files

At this time, auxiliary summary tables and the full validation report are not distributed as separate files in this repository. Key validation findings and literature sources are summarized in this dataset card.

---

7. Generation Methodology (Summary)

The dataset was generated using an internal, literature-driven synthetic data framework, following a structured multi-phase pipeline:

1. Domain specification – Define gene panels (Oncotype DX, PAM50, Ki-67) and African focus.
2. Literature review – Extract subtype frequencies, Ki-67 levels, and score distributions from 30+ studies.
3. Parameter configuration – Encode all parameters in YAML (gene_expression_config.yaml).
4. Generation – Use a reproducible Python generator to simulate 50,000 cases.
5. Validation – Run scripts/validate_gene_expression.py to check distributions, coherence, and correlations (30 checks).
6. Documentation – Create dataset card and usage examples.
7. Release – Upload to Hugging Face with CC-BY-NC-4.0.

7.1 Key Modeling Choices

• PAM50 subtypes assigned using population-specific target proportions.
• Receptor status inferred from subtypes with high concordance (e.g., Basal_like ↔ TNBC).
• Gene expression values drawn from subtype-specific normal distributions calibrated to literature and TCGA-like ranges.
• Ki-67 modeled as a function of subtype, population multiplier, and MKI67 expression.
• Oncotype DX scores simulated for ER+/HER2− tumors using population and subtype–dependent distributions.

---

8. Validation Status

• This dataset has been validated using scripts/validate_gene_expression.py on the released CSV.

• ✅ Generation completed for 50,000 samples.

• ✅ 30 validation checks executed (structure, distributions, coherence).

• ✅ Results: 25 PASS, 5 WARN, 0 FAIL.

• ✅ Validation report: maintained internally; main findings are summarized below.

Key validation themes include:

• Distributions vs literature (subtypes, TNBC, Ki-67, Oncotype)
• Expression ranges and subtype-specific patterns
• ER/ESR1 and HER2/ERBB2 concordance
• Basal-like ↔ TNBC overlap
• Ki-67 correlation with MKI67 expression and grade

---

9. Example Usage

9.1 Load with pandas

import pandas as pd

df = pd.read_csv("gene_expression_data.csv")
print(df.shape)
print(df.head())

9.2 Load with datasets

from datasets import load_dataset

dataset = load_dataset("electricsheepafrica/tumor-gene-expression-african")
df = dataset["train"].to_pandas()

print(f"Samples: {len(df)}")
print(f"Columns: {len(df.columns)}")

9.3 Simple Analysis: PAM50 Subtypes

subtype_counts = df["pam50_subtype"].value_counts(normalize=True) * 100
print(subtype_counts)

9.4 Ki-67 by Subtype

ki67_by_subtype = df.groupby("pam50_subtype")["ki67_percentage"].describe()
print(ki67_by_subtype)

9.5 Oncotype DX Risk in ER+/HER2− Tumors

mask = (df["ER_status"] == "Positive") & (df["HER2_status"] == "Negative")
rs = df.loc[mask, "oncotype_RS"]

print(rs.describe())
print(df.loc[mask, "oncotype_risk_category"].value_counts(normalize=True) * 100)

---

10. Ethical & Appropriate Use

10.1 Appropriate Uses ✅

• Method development and benchmarking
• Educational demos and tutorials
• Health equity and disparity analysis (synthetic)
• Robustness testing for molecular classifiers

10.2 Inappropriate Uses ❌

• Clinical decision-making for individual patients
• Real-world prognosis or treatment selection
• Insurance, employment, or financial decisions
• Genetic ancestry inference at the individual level

This dataset is synthetic and must not be used as a substitute for real clinical or genomic data in patient care.

---

11. License

• License: CC-BY-NC-4.0
• Commercial use: Not permitted without explicit permission.

If you use this dataset in academic or non-commercial work, please cite as below.

---

12. Citation

Electric Sheep Africa (2025).
Tumor Gene Expression Panels in African Breast Cancer (Synthetic Dataset).
Generated using an internal, literature-driven synthetic data methodology.
Hugging Face Datasets. Version 1.0.0.

---

13. Contact

• Organization: Electric Sheep Africa
• Hugging Face: https://huggingface.co/electricsheepafrica

Feedback and collaboration inquiries are welcome.