README.md

---
license: cc-by-nc-4.0
task_categories:
- tabular-classification
- tabular-regression
language:
- en
tags:
- synthetic
- healthcare
- genomics
- variant-calling
- vcf
- vep
- cadd
- clinvar
- gnomad
- rna-seq
- bulk-rna-seq
- single-cell
- scrna-seq
- pbmc
- gene-expression
- polygenic-risk-score
- prs
- pharmacogenomics
- pgx
- cpic
- pharmgkb
- cyp2d6
- cyp2c19
- ancestry
- grch38
- hg38
- titv
- hardy-weinberg
- tabula-sapiens
- gtex
- 10x-genomics
pretty_name: HLT-013 Synthetic Multi-Modal Genomics Dataset (Sample Preview)
size_categories:
- 1K<n<10K
---

# HLT-013 — Synthetic Multi-Modal Genomics Dataset (Sample Preview)

**A free, schema-identical preview of the full HLT-013 commercial product from [XpertSystems.ai](https://xpertsystems.ai).**

A **fully synthetic** multi-modal genomics dataset combining **variant calls** (VCF-style with VEP/CADD/ClinVar/gnomAD annotations), **bulk RNA-seq gene expression** (5 tissues × 2,000 genes), **single-cell RNA-seq PBMC profiles** (10 cell types), **polygenic risk scores** (50 traits across 5 disease domains), and **pharmacogenomics star allele calls** (25 CPIC-actionable genes) — all linked through 1,000 individuals across 5 ancestry superpopulations (EUR/AFR/EAS/AMR/SAS, gnomAD-calibrated).

> ⚠️ **PRIVACY & SYNTHETIC NATURE**
> Every record in this dataset is **100% synthetic**. **No real patient data, no PHI, no real genome sequences, no real variant calls.** Population-level distributions match published gnomAD / ClinVar / VEP / CPIC / Tabula Sapiens benchmarks but the genomic profiles are computationally generated.

---

## What's in this sample

**1,000 individuals × 6 multi-modal genomic data tables linked by `sample_id`.**

| File | Rows × Cols | Description |
|---|---|---|
| `cohort_manifest.csv` | 1,000 × 10 | Individual master — ancestry, sex, age, sequencing type, mean coverage, pct bases ≥20x, consent tier |
| `variants_annotated.csv` | 600 × 14 | VCF-style: CHROM/POS/RSID/REF/ALT/GT/GQ/DP + VEP consequence + CADD Phred + ClinVar sig + gnomAD AF + HWE p-value |
| `gene_expression.csv` | 2,000 × 7 | Bulk RNA-seq gene panel — mean log2TPM, SD, CV, % expressed, housekeeping flag |
| `scrna_pbmc.csv` | ~1,900 × 11 | Single-cell PBMC: cell type, cluster ID, UMAP coords, n_genes, n_counts, pct_mito, doublet score |
| `polygenic_risk_scores.csv` | 1,000 × 152 | 50 PRS traits × (raw score + ancestry-adjusted percentile + risk tier) per individual |
| `pharmacogenomics.csv` | 1,000 × 102 | 25 PGx genes × (star allele class + CPIC recommendation + ACMG actionable + drug-specific guidance) |
| `metadata.json` | — | Run manifest: seed, genome build, ancestry distribution, Ti/Tv, tissue/cell-type/PRS/PGx counts |

**Total:** ~2.4 MB across 8 files.

---

## Schema highlights

### `cohort_manifest.csv` (10 columns)

`sample_id`, `ancestry_superpop` (EUR/AFR/EAS/AMR/SAS), `sex`, `age`, `cohort_id`, `genome_build` (GRCh38), `sequencing_type` (WGS/WES), `mean_coverage` (numeric, 30x WGS / 100x WES mix), `pct_bases_20x` (QC metric), `consent_tier` (research/clinical/broad)

### `variants_annotated.csv` (14 columns)

**Position:** `CHROM` (1-22, X), `POS` (genomic coordinate), `RSID` (rs identifier), `REF`, `ALT`, `variant_type` (SNP/InDel)

**Genotype:** `GT` (0/0, 0/1, 1/1), `GQ` (genotype quality 0-99), `DP` (read depth)

**Annotation:** `consequence` (VEP terms: intergenic / intron / synonymous / missense / 3'UTR / 5'UTR / splice_region / stop_gained / splice_donor / splice_acceptor / frameshift / inframe_indel), `CADD_phred` (deleteriousness 0-50+), `ClinVar_sig` (Benign / Likely_benign / VUS / Likely_pathogenic / Pathogenic), `AF_gnomAD` (allele frequency 0-1)

**Population genetics:** `HWE_pval` (Hardy-Weinberg Equilibrium p-value)

### `gene_expression.csv` (7 columns)

`gene_id` (ENSG-style), `gene_name`, `mean_log2TPM`, `sd_log2TPM`, `cv` (coefficient of variation), `pct_expressed` (% of individuals with detectable expression), `is_housekeeping` (flag for stably-expressed reference genes)

### `scrna_pbmc.csv` (11 columns)

`sample_id`, `cell_barcode`, `cell_type` (10 types: CD4_T_naive, CD4_T_memory, CD8_T_cytotoxic, B_cell_naive, B_cell_memory, NK_cell, Monocyte_classical, Monocyte_nonclassical, pDC, Platelet), `cluster_id`, `UMAP_1`, `UMAP_2`, `n_genes`, `n_counts`, `pct_mito`, `doublet_score`, `cell_type_confidence`

### `polygenic_risk_scores.csv` (152 columns)

`sample_id`, `ancestry`, plus **50 traits × 3 fields each**:

- `PRS_<trait>` — raw polygenic score
- `PRS_<trait>_pct` — ancestry-adjusted percentile (0-100)
- `PRS_<trait>_tier` — risk tier (Low/Intermediate/High)

**50 traits across 5 domains:**
- **CVD (10):** coronary_artery_disease, atrial_fibrillation, heart_failure, hypertension, stroke, peripheral_artery_disease, abdominal_aortic_aneurysm, hypertrophic_cardiomyopathy, dilated_cardiomyopathy, long_qt_syndrome
- **Metabolic (10):** type2_diabetes, BMI, LDL_cholesterol, HDL_cholesterol, triglycerides, fasting_glucose, HbA1c, type1_diabetes, obesity, NAFLD
- **Oncology (10):** breast_cancer, prostate_cancer, colorectal_cancer, lung_cancer, melanoma, ovarian_cancer, pancreatic_cancer, glioma, leukemia, lymphoma
- **Autoimmune (10):** rheumatoid_arthritis, type1_diabetes_autoimmune, MS, lupus, psoriasis, IBD_crohns, IBD_ulcerative_colitis, celiac, asthma, atopic_dermatitis
- **Psychiatric (10):** depression, bipolar, schizophrenia, anxiety, ADHD, autism, alzheimers, parkinsons, alcohol_dependence, smoking_behavior

### `pharmacogenomics.csv` (102 columns)

`sample_id`, `ancestry`, plus **25 genes × 4 fields each**:

- `PGx_<gene>_class` — predicted phenotype (NM=Normal Metabolizer / IM=Intermediate / PM=Poor / RM=Rapid / UM=Ultrarapid)
- `PGx_<gene>_CPIC` — CPIC dosing recommendation (Standard/Reduce/Increase/Avoid)
- `PGx_<gene>_ACMG_actionable` — ACMG SF v3.2 actionable flag
- `PGx_<gene>_recommendation` — drug-specific guidance text

**25 genes (CPIC Level A or B):** CYP2D6, CYP2C19, CYP2C9, CYP3A5, CYP2B6, CYP1A2, TPMT, NUDT15, DPYD, SLCO1B1, VKORC1, UGT1A1, HLA-B, HLA-A, CYP4F2, IFNL3, G6PD, RYR1, CACNA1S, ABCG2, F5, MTHFR, NAT2, ATM, BRCA1

---

## Calibration source story

The full HLT-013 generator anchors all distributions to authoritative genomics references:

- **gnomAD v4 (Karczewski et al. 2020)** — Population allele frequencies, ancestry superpopulation proportions
- **VEP (McLaren et al. 2016)** — Variant Effect Predictor consequence annotation
- **ClinVar (Landrum et al. 2018)** — Clinical variant significance database
- **CADD (Rentzsch et al. 2019)** — Combined Annotation Dependent Depletion scores
- **CPIC** — Clinical Pharmacogenetics Implementation Consortium dosing guidelines
- **PharmGKB** — Gene-drug interaction knowledge base
- **10x Genomics PBMC 10k reference** — Single-cell PBMC cell type proportions
- **Tabula Sapiens (Quake Lab)** — Cross-tissue cell type catalog
- **GTEx Consortium** — Tissue-specific gene expression
- **PGS Catalog (Lambert et al. 2021)** — Polygenic score trait coverage
- **ACMG SF v3.2** — Secondary findings actionable variant list

### Sample-scale validation scorecard

| Metric | Observed | Target | Status | Source |
|---|---|---|---|---|
| EUR ancestry share | 40.2% | 40% ± 5% | ✅ PASS | gnomAD v4 |
| Ancestry superpop count | 5 | 5 | ✅ PASS | gnomAD |
| Ti/Tv ratio | 2.31 | 2.06 ± 0.80 | ✅ PASS | Wang et al. (2015) |
| ClinVar P/LP rate | 2.3% | 2.5% ± 2.5% | ✅ PASS | ClinVar |
| Mean coverage | 42.1x | 42 ± 10 | ✅ PASS | Clinical genomics QC |
| Cell type diversity | 10 | 10 | ✅ PASS | 10x Genomics PBMC 10k |
| PRS trait count | 50 | 50 | ✅ PASS | PGS Catalog |
| PGx gene count | 25 | 25 | ✅ PASS | CPIC Level A/B |
| CYP2D6 NM rate | 66.7% | 65% ± 15% | ✅ PASS | CPIC + PharmGKB |
| Expression tissue count | 5 | 5 | ✅ PASS | Tabula Sapiens / GTEx |

**Grade: A+ (100/100) — verified across 6 random seeds (42, 7, 123, 2024, 99, 1).**

---

## Loading examples

### Pandas — explore the cohort

```python
import pandas as pd

cohort = pd.read_csv("cohort_manifest.csv")
variants = pd.read_csv("variants_annotated.csv")
prs = pd.read_csv("polygenic_risk_scores.csv")
pgx = pd.read_csv("pharmacogenomics.csv")

# Ancestry distribution
print(cohort["ancestry_superpop"].value_counts(normalize=True).round(3))

# Variant consequence breakdown
print(variants["consequence"].value_counts(normalize=True).round(3))

# ClinVar significance
print(variants["ClinVar_sig"].value_counts())
```

### Variant filtering

```python
import pandas as pd

variants = pd.read_csv("variants_annotated.csv")

# High-impact variants (Pathogenic + CADD > 25)
high_impact = variants[
    (variants["ClinVar_sig"].isin(["Pathogenic", "Likely_pathogenic"])) |
    (variants["CADD_phred"] > 25)
]
print(f"High-impact variants: {len(high_impact)}")

# Rare variants (gnomAD AF < 1%)
rare = variants[variants["AF_gnomAD"] < 0.01]
print(f"Rare variants: {len(rare)}")

# HWE-departure variants
hwe_violations = variants[variants["HWE_pval"] < 0.001]
print(f"HWE violations: {len(hwe_violations)}")
```

### PRS risk stratification

```python
import pandas as pd

prs = pd.read_csv("polygenic_risk_scores.csv")

# Top 10% CAD risk individuals
high_cad = prs[prs["PRS_coronary_artery_disease_pct"] >= 90]
print(f"High CAD risk: {len(high_cad)} individuals")

# Multi-trait risk profile
risk_traits = ["coronary_artery_disease", "type2_diabetes", "breast_cancer"]
for trait in risk_traits:
    tier_col = f"PRS_{trait}_tier"
    if tier_col in prs.columns:
        print(f"\n{trait} risk tier distribution:")
        print(prs[tier_col].value_counts(normalize=True).round(3))
```

### PGx phenotype distribution

```python
import pandas as pd

pgx = pd.read_csv("pharmacogenomics.csv")

# CYP2D6 phenotype by ancestry
print(pd.crosstab(pgx["ancestry"], pgx["PGx_CYP2D6_class"], normalize="index").round(3))

# Actionable findings (ACMG SF v3.2)
actionable_cols = [c for c in pgx.columns if c.endswith("_ACMG_actionable")]
n_actionable = pgx[actionable_cols].sum(axis=1)
print(f"\nIndividuals with ≥1 actionable PGx finding:")
print(f"  None:    {(n_actionable == 0).sum()}")
print(f"  1 gene:  {(n_actionable == 1).sum()}")
print(f"  2+ genes: {(n_actionable >= 2).sum()}")
```

### scRNA-seq cell type analysis

```python
import pandas as pd

scrna = pd.read_csv("scrna_pbmc.csv")

# Cell type proportions per sample
cell_pcts = scrna.groupby("sample_id")["cell_type"].value_counts(normalize=True).unstack(fill_value=0)
print("Mean cell type proportions:")
print(cell_pcts.mean().sort_values(ascending=False).round(3))

# QC metrics by cell type
print("\nQC metrics by cell type:")
print(scrna.groupby("cell_type")[["n_genes", "pct_mito", "doublet_score"]].mean().round(2))
```

### Hugging Face Datasets

```python
from datasets import load_dataset

ds = load_dataset("xpertsystems/hlt013-sample", data_files={
    "cohort":     "cohort_manifest.csv",
    "variants":   "variants_annotated.csv",
    "expression": "gene_expression.csv",
    "scrna":      "scrna_pbmc.csv",
    "prs":        "polygenic_risk_scores.csv",
    "pgx":        "pharmacogenomics.csv",
})
print(ds)
```

---

## Suggested use cases

- **Variant prioritization ML** — train classifiers on CADD + ClinVar + AF features to predict pathogenicity
- **PRS-disease prediction modeling** — multi-trait ML for absolute risk stratification
- **Ancestry imputation** — train ancestry callers from variant features
- **Variant Effect Predictor pipeline testing** — schema-compliant data for VEP/SnpEff annotation pipeline development
- **Pharmacogenomic CDS rules engine testing** — populate PGx clinical decision support systems
- **scRNA-seq cell type classification** — train cell type callers from gene expression + UMAP coordinates
- **HWE violation detection** — flag spurious genotype calls or population structure
- **Multi-modal genomics integration** — joint modeling across variants + expression + PRS + PGx
- **Clinical genomics LIMS testing** — populate clinical genomics pipelines with realistic synthetic patients
- **Healthcare AI pretraining** — pretrain models on synthetic genomic profiles before fine-tuning on real biobank data
- **Educational use** — graduate genomics, biostatistics, and precision medicine coursework

---

## Sample vs. full product

| Aspect | This sample | Full HLT-013 product |
|---|---|---|
| Individuals | 1,000 | 100,000+ (default) up to 1M |
| Variants per individual | 600 representative | Full WGS ~6.5M variants |
| Genes (bulk expression) | 2,000 | Full transcriptome ~20,000 genes |
| scRNA cells per sample | ~2 (sampled) | ~200 cells per sample |
| PRS traits | 50 | 50 (full coverage) |
| PGx genes | 25 | 25 (full CPIC Level A/B coverage) |
| Schema | identical | identical |
| Calibration | identical | identical |
| License | CC-BY-NC-4.0 | Commercial license |

The full product unlocks:
- **Up to 1M individuals** for biobank-scale genomic ML training
- **Full WGS variant calls** (~6.5M variants per individual)
- **Full transcriptome** (20,000+ genes)
- **Dense scRNA-seq profiles** (200+ cells per sample)
- **GWAS summary statistics** for the 50 PRS traits
- **Family pedigrees** for trio/quartet analysis
- Commercial use rights

**Contact us for the full product.**

---

## Limitations & honest disclosures

- **Sample is preview-only.** 1,000 individuals × 600 representative variants is enough to demonstrate schema and calibration, but is **not statistically sufficient** for serious GWAS, PRS development, or rare variant analysis. Use the full product (100K+) for serious work.
- **Variant set is sub-sampled.** Each individual carries 600 representative variants (mix of SNPs + InDels), not the ~6.5M variants from full WGS. Variant positions are real-coordinate-valid but sparse.
- **scRNA-seq cells per sample are sparse (~2 cells/sample at preview scale).** Real PBMC scRNA-seq experiments yield 200-1000 cells per sample. The sample compresses this for size — full product has dense per-sample profiles.
- **Gene expression is panel-summary, not per-individual.** The `gene_expression.csv` file gives population-level summary statistics (mean log2TPM, SD, CV) across the 1,000 individuals, NOT individual-specific TPM values. For per-individual expression matrices, use the full product.
- **Housekeeping gene flag rate runs slightly high (~7.5% vs typical 1-3%).** The generator marks more genes as housekeeping than strict biological definitions. Cross-reference with HK genes lists (Eisenberg & Levanon 2013) if exact housekeeping calls matter.
- **Ti/Tv ratio variance is high at 600-variant sample scale (1.78-2.68 across seeds vs target 2.06).** This is small-sample noise — full WGS at 6.5M variants converges tightly to the gnomAD target.
- **RSIDs are synthetic.** Generated RSIDs follow the rsXXXXXXX format but do NOT correspond to real dbSNP entries.
- **gnomAD AF values are sampled from realistic distributions but are NOT real allele frequencies.** Do not use this data for variant frequency reporting.
- **ClinVar IDs not included.** Variants have `ClinVar_sig` classifications but no real ClinVar variation IDs.
- **PRS scores are simulated, not based on real GWAS effect sizes.** Distributions match published PRS percentile shapes but specific scores do NOT reflect real allele effects.
- **PGx phenotype calls follow CPIC frequency distributions but are NOT mechanistic.** Star allele class assignments are population-frequency-driven, not derived from underlying CYP/TPMT/etc. variant calls in `variants_annotated.csv`.
- **Synthetic, not derived from real biobank data.** Distributions match published gnomAD/ClinVar/CPIC/Tabula Sapiens references but do NOT reflect any specific real cohort (UK Biobank, All of Us, etc.).

---

## Ethical use guidance

This dataset is designed for:
- Genomic ML methodology development
- Clinical genomics pipeline testing
- PRS modeling research
- Pharmacogenomics CDS rule engine development
- scRNA-seq cell type annotation methodology
- Healthcare AI pretraining for genomic prediction tasks
- Educational use in clinical genomics, precision medicine, and biostatistics

This dataset is **not appropriate for**:
- Making clinical genetic diagnoses about real individuals
- Real PRS reporting for real patients without validated ancestry-matched reference panels
- Pharmacogenomic prescribing decisions for real patients without CPIC consultation
- Variant pathogenicity calls without ACMG framework validation on real ClinVar data
- Ancestry-based discriminatory modeling
- Population-genetic claims about real ethnic groups

---

## Companion datasets in the Healthcare vertical

- [HLT-001](https://huggingface.co/datasets/xpertsystems/hlt001-sample) — Synthetic Patient Population (CDC/NHANES)
- [HLT-002](https://huggingface.co/datasets/xpertsystems/hlt002-sample) — Synthetic EHR (FHIR R4)
- [HLT-003](https://huggingface.co/datasets/xpertsystems/hlt003-sample) — Synthetic Clinical Trial
- [HLT-004](https://huggingface.co/datasets/xpertsystems/hlt004-sample) — Synthetic Disease Progression
- [HLT-005](https://huggingface.co/datasets/xpertsystems/hlt005-sample) — Synthetic Hospital Admission
- [HLT-006](https://huggingface.co/datasets/xpertsystems/hlt006-sample) — Synthetic Medical Imaging
- [HLT-007](https://huggingface.co/datasets/xpertsystems/hlt007-sample) — Synthetic Drug Response
- [HLT-008](https://huggingface.co/datasets/xpertsystems/hlt008-sample) — Synthetic Healthcare Claims
- [HLT-009](https://huggingface.co/datasets/xpertsystems/hlt009-sample) — Synthetic ICU Vital Sign Monitoring
- [HLT-010](https://huggingface.co/datasets/xpertsystems/hlt010-sample) — Synthetic Hospital Resource Usage
- [HLT-011](https://huggingface.co/datasets/xpertsystems/hlt011-sample) — Synthetic Rare Disease + Trial Eligibility
- [HLT-012](https://huggingface.co/datasets/xpertsystems/hlt012-sample) — Synthetic Pandemic Spread
- **HLT-013** — Synthetic Multi-Modal Genomics (you are here)

Use **HLT-001 through HLT-013 together** for the full healthcare data stack — and HLT-013 specifically extends the catalog into **precision medicine & clinical genomics**, complementing HLT-007 (drug response with PGx hooks) and HLT-011 (rare disease with gene-variant calls).

---

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{xpertsystems_hlt013_sample_2026,
  author       = {XpertSystems.ai},
  title        = {HLT-013 Synthetic Multi-Modal Genomics Dataset (Sample Preview)},
  year         = 2026,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/xpertsystems/hlt013-sample}
}
```

---

## Contact

- **Web:** [https://xpertsystems.ai](https://xpertsystems.ai)
- **Email:** [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai)
- **Full product catalog:** Cybersecurity, Insurance & Risk, Materials & Energy, Oil & Gas, Healthcare, and more

**Sample License:** CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0)
**Full product License:** Commercial — please contact for pricing.