Generative_Genomic_Data

Synthetic genotype generation for admixed Latin American cohorts using a conditional Wasserstein GAN with gradient penalty (cWGAN-GP) trained in principal-component space and conditioned on continental ancestry proportions (AFR / EUR / AMR).

The goal is to generate genotypes for a target population given its known admixture profile, so that synthetic individuals reproduce both the genome-wide correlation structure and the ancestry composition of the real cohort.

Method

The generator never sees raw genotypes. Everything happens in a compressed PC space fit on a reference panel:

1. Load target-cohort and reference genotype tensors (additive dosage, 0/1/2).
2. Intersect to the SNPs shared by both; impute residual missingness to per-SNP means.
3. Fit PCA on the reference panel (default 50 PCs) and project both datasets in.
4. Assign admixture conditions. Reference samples get their known one-hot label ([1,0,0] / [0,1,0] / [0,0,1]); target samples get a barycentric projection onto the triangle formed by the AFR/EUR/AMR reference centroids.
5. Train the cWGAN-GP jointly on reference + target samples, with both generator and critic conditioned on the 3-dim admixture vector.
6. Sample ancestry proportions from a Dirichlet fit to the observed target admixture (or pass --target_admixture explicitly).
7. Invert the PCA, discretize back to 0/1/2, and write genotypes plus a PCA overlay plot.

The model drives generation end to end — there is no post-hoc Dirichlet resampling of the GAN output.

Repository layout

src/
  conditional_wgan_pipeline.py   # the cWGAN-GP model — the only model script here
  plot_colors.py                 # per-population colors; imported by the pipeline
  vcf_to_unphased_tensor.py      # VCF -> additive dosage tensor (builds pipeline input)
docs/
  pipeline_walkthrough.md        # line-by-line walkthrough of the cWGAN pipeline
results/<COHORT>/
  cWGAN_PCA/                     # PCA overlay plots (committed)
  cWGAN_gen-data/                # synthetic genotype CSVs (gitignored — large)

src/ is deliberately scoped to the cWGAN pipeline. Earlier work — the unconditional PCA+WGAN pipeline, the VAE experiments, and the standalone PCA and evaluation scripts — is kept out of the repo via .gitignore rather than deleted, so it still exists in the original working tree.

data/, snps/, and slurm/ are gitignored: they hold raw VCFs, PLINK binaries, genotype tensors, SNP lists, and job logs. Synthetic genotype CSVs are also gitignored — they are large and fully reproducible from the pipeline.

Running

python src/conditional_wgan_pipeline.py \
    --data_input     data/<COHORT>/<COHORT>_refintersect_qc_ldpruned_unphased_tensor.csv \
    --ref_input      data/reference/extractedChrAll_<COHORT>_intersect_qc_ldpruned_unphased_tensor.csv \
    --fam            data/reference/extractedChrAll_common_snps_forReference.fam \
    --output         results/<COHORT>/cWGAN_gen-data/synthetic_genotypes_seed42.csv \
    --plot           results/<COHORT>/cWGAN_PCA/pca_overlay_seed42.png \
    --sample_name    <COHORT> \
    --n_components   50 \
    --n_synthetic    200 \
    --wgan_epochs    1000 \
    --wgan_latent_dim 32 \
    --batch_size     64 \
    --seed           42

To generate for an explicit ancestry profile instead of the cohort-fitted Dirichlet, add --target_admixture 0.6,0.2,0.2 (AFR, EUR, AMR).

Other useful flags: --alpha_concentration (Dirichlet tightness), --variance_scale, --real_color.

Environment

environment.yml defines the full conda environment (GGD_env). Runs on PACE require a GPU node; the pipeline falls back to CPU but is substantially slower.

conda env create -f environment.yml && conda activate GGD_env

Current results

Latest sweep: 4 cohorts x 5 seeds (42-46), 20 runs, all completed. Reference panel: 644 samples (AFR = 312, EUR = 273, AMR = 59). 50 PCs, 1000 epochs, 200 synthetic samples per run. PCA overlays are in results/<COHORT>/cWGAN_PCA/.

Cohort	Real N	SNPs (post-intersect)	Missingness imputed	Var. explained (50 PC)	Fixed SNPs	Assessment
BOG	337	28,514	0.16%	20.1%	0–1	Usable
UKB	245	77,387	0.31%	19.1%	46–58	Usable
VDC3	24	6,688	47.6%	23.9%	1–2	Over-dispersed
SAI	10	50,795	72.7%	24.7%	107–150	Not usable

Known limitations

• Small-N cohorts fail. SAI (10 samples, 72.7% imputed) produces a degenerate fit: the Dirichlet collapses to alpha = [261.25, 0.00, 47.95], giving an exactly zero EUR component and effectively no variance in the sampled conditions. The synthetic cloud spans far beyond the real data and partly outside the ancestry triangle. VDC3 (24 samples, 47.6% imputed) is over-dispersed for the same reason. Treat both as diagnostic, not as usable synthetic data.
• Mode-dropping on the admixed tail. For BOG and UKB the synthetic distribution covers the core of the real cloud but under-represents individuals extending toward the AFR vertex.
• Low retained variance. 50 PCs capture only 19–25% of total variance, so reconstruction discards most of the fine-scale signal by construction.
• Stale log labels. The pipeline prints BAQ for every cohort (e.g. "Loading BAQ genotypes...") regardless of --sample_name. Cosmetic only — the data loaded is correct — but it makes logs confusing to read.