PCA-WGAN

A phased haplotype over a million SNPs is a million-dimensional binary vector, and a generator whose final dense layer maps even a 1,000-unit latent onto it carries a billion weights before it learns anything about genetics. That arithmetic is what has kept artificial-genome models confined to short windows: fully connected architectures that worked on a few hundred informative markers do not survive the jump to genome scale.

PCA-WGAN leaves the SNP space entirely. Principal component analysis is fitted once on the training panel, a Wasserstein GAN with gradient penalty learns the distribution of the resulting principal-component scores, and generation runs the other way: sample the generator from Gaussian noise, inverse-transform the synthetic scores through the real-data eigenvectors, binarize at 0.5. PCA earns its place because it inverts exactly, unlike t-SNE or UMAP, and costs nothing to fit, unlike an autoencoder; the GAN keeps the model nonlinear despite the linear compression. Reconstructing held-out real haplotypes through the training basis misses only 6% of sites, well below the smallest genetic distance separating two individuals in the panel.

The framework comes from Antoine Szatkownik, Cyril Furtlehner, Guillaume Charpiat, Burak Yelmen and Flora Jay at Université Paris-Saclay (CNRS, Inria, LISN) with the University of Tartu, posted to bioRxiv in August 2024. The same group later swapped the GAN generator for a conditional diffusion model in Light PCA-DDPM; this is the earlier of the two.

#Key Features

• Generation in principal-component space: The generative network never sees a SNP. It learns the distribution of PC scores, where ancestry-linked variation is concentrated, and the inverse transform reconstructs the haplotype.
• Glocal-PCA-WGAN: A chained variant that splits the SNP matrix into three equal blocks, fits an independent PCA to each, and trains local critics on the blocks alongside one global critic on their concatenation.
• Light PCA-WGAN: A frugal variant that gives the GAN only the first six PCs — the ones with multi-modal structure — and draws the remaining 5,002 from a multivariate normal whose parameters are fixed or predicted from those six by a small network.
• Shrinkage correction: PC scores of generated samples are under-dispersed relative to the training set, distorting nearest-neighbour privacy metrics; rescaling each PC by the inverse empirical shrinkage factor restores the resemblance score.
• A motif-diversity metric: Counting unique k-SNP motifs in non-overlapping windows gives an information-theoretic handle on local haplotypic structure that summary statistics miss.

#Technical Details

Training uses the 1000 Genomes panel curated by Yelmen et al. — 2,504 individuals, 5,008 phased haplotypes, 65,535 contiguous SNPs spanning chr1:534247–81813279 (about 80 Mb of the Omni 2.5 array framework) across 26 worldwide populations. PCA retains 4,507 components, 90% of the PCs. On a single A100-40GB, PCA-WGAN carries 74M parameters and trains in 20 hours and Glocal-PCA-WGAN 140M in two days, while Light PCA-WGAN needs 58K parameters with a fixed multivariate normal or 750K with the predictive network and about two hours either way — against 16M parameters and six to eight days for the prior convolutional WGAN. PCA-WGAN's linkage-disequilibrium decay curve is the closest to real data of all models compared. Nearest-neighbour adversarial accuracy needed the most work: PCA-WGAN and Glocal-PCA-WGAN land near the ideal value of 0.5 only because their two components are badly imbalanced, and the shrinkage correction lifts the honest score above the convolutional WGAN's and the RBM's at the cost of slightly worse LD, while Light PCA-WGAN with NN scores best of all. The clearest quality cost is allele fixing: PCA-WGAN drives 20,523 sites (roughly 31%) to a single allele, against 12,686 for Glocal-PCA-WGAN, 2,672 for the convolutional WGAN and 8 in real data.

#Applications

The output is a proxy dataset — synthetic haplotypes a data holder can release in place of genotypes bound by consent terms and biobank access agreements, for genome-wide association studies, local ancestry inference, imputation and selection scans. The compute profile matters as much as the quality: a lab with one GPU can train Light PCA-WGAN in two hours and still get population structure, allele frequencies and short-range LD that track the real panel.

#Impact

The contribution is a recipe rather than a checkpoint: compress with a cheap invertible transform, generate in the reduced space, invert. It makes fully connected generators affordable at genome scale, and it transfers to other generative families. The costs are stated plainly: allele fixing at roughly a third of sites, a synthetic k-SNP motif diversity higher than real genomes at every window size tested, and a fixed 65,535-SNP window on chromosome 1 with binary allele calls. This bioRxiv version has not been peer reviewed. Training code for all three variants and the real training matrix are public in the group's Inria GitLab repository, which carries no license and no trained checkpoints.