Loading…
Frugal conditional diffusion model that samples human SNP haplotypes in PCA space, producing artificial genomes for a chosen continental ancestry.
No providers recorded yet. Browse all providers
A reference panel for local ancestry inference is a list of real people's haplotypes, which is precisely what makes it hard to share: consent terms and biobank access agreements sit between a researcher and the populations least represented in public genomic archives. Artificial genomes — synthetic haplotypes drawn from a generative model rather than an individual — are the way around that, but the dimensionality is punishing. A haplotype over 65,535 SNPs is a 65,535-dimensional binary vector, and one dense layer mapping even a 1,000-unit latent onto it costs 65 million weights.
Light PCA-DDPM argues that almost none of that capacity is needed. Principal component analysis of a SNP panel concentrates continental population structure into a handful of leading axes, and PCA inverts cheaply, so the model runs a denoising diffusion probabilistic model over the first six principal components only. The remaining 2,134 low-variance components are drawn from a diagonal multivariate normal fitted to the real PC scores; the two halves are concatenated, inverse-transformed to SNP space and binarized at 0.5. The denoising network is conditioned on a continental super-population label, so generation consumes Gaussian noise plus a label and nothing else — no real individual's genotype enters the sampling path.
The model was built at Université Paris-Saclay with CNRS, Inria and the LISN lab, alongside collaborators at UNAM, Brown University, Trinity College Dublin and the University of Tartu, and posted as a bioRxiv preprint in October 2024. It succeeds the same group's Light PCA-WGAN, with the Wasserstein GAN generator swapped for a conditional DDPM.
The denoiser is a small multilayer perceptron: a sinusoidal time embedding, a three-dimensional learned label embedding concatenated to the six input features, 256-unit linear layers, and six outputs. Diffusion uses 1,000 steps with a linear noise schedule from 0.0001 to 0.02 and the standard noise-prediction MSE objective. Training data is 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, coded against GRCh37). Quality experiments trained on the full panel across five super-populations; the ancestry-inference experiments used a three-continent subset of 1,188 unadmixed individuals balanced across AFR, EAS and EUR.
On summary statistics — PCA structure, allele frequency spectra, linkage disequilibrium and three-point correlations — Light PCA-DDPM performs on par with Light PCA-WGAN, and both capture LD that a per-site Bernoulli baseline cannot. For utility, LAI-Net trained on real haplotypes reached 94.84% ± 0.1 weighted accuracy on simulated admixed test data; trained instead on Light PCA-DDPM output it reached 94.16–94.43%, while the Bernoulli baseline fell roughly 10% behind. On 504 real admixed American samples, ancestry calls from a classifier trained on synthetic data agreed with one trained on real data across 93% of the genome, against 83% for the Bernoulli baseline.
The intended use is supplying ancestry reference panels where real ones are restricted: a data holder trains the model once and shares synthetic haplotypes instead of genotypes. The strongest practical result is augmentation — when only a small number of real samples are available, adding 8,554 synthetic haplotypes recovers much of the accuracy the shortfall costs, which matters most for the underrepresented groups whose panels are thinnest. Training several models under different seeds and averaging their downstream predictions, a deep generative ensemble, gave only marginal further gains.
Quality and utility come apart here: the Bernoulli baseline produces genomes poor by every population-genetics statistic yet still trains a usable ancestry classifier, which argues for judging synthetic genomic data against the task it will serve rather than a general realism threshold. The model's scope is deliberately narrow — one fixed 65,535-SNP window on chromosome 1, continental-scale labels rather than the 26 individual 1000 Genomes populations, and binary allele calls at a fixed threshold — and conditioning on smaller, sparser modes remains unsolved. The work is a preprint and has not been peer reviewed. Training and generation code is public in the group's Inria GitLab repository, as is the training matrix, but the repository carries no license and no trained checkpoints.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.