Genetic Probabilistic Circuits offer a tractable deep generative model for population genomic data

A team including Prateek Anand, Anji Liu, Meihua Dang, Boyang Fu, Xinzhu Wei, Guy Van den Broeck, and Sriram Sankararaman has published a study in PLOS Genetics introducing Genetic Probabilistic Circuits (GPC), a new class of deep generative model designed specifically for genetic variation data.

Generative models of genotype data serve several functions in population genetics and statistical genomics: they are used to simulate artificial genomes for benchmarking methods and testing evolutionary hypotheses, and to construct reference panels for genotype imputation — allowing researchers to work with privacy-compliant synthetic data rather than sharing real participant records.

Existing approaches face a tension between expressiveness and tractability. Models that capture complex dependencies across the genome tend to be computationally demanding and difficult to interpret or scale; simpler models sacrifice fidelity to the true distribution of genetic variation. The GPC framework is described by the authors as addressing both limitations simultaneously, using a circuit-based probabilistic architecture that allows efficient exact inference while modelling higher-order dependencies among genetic variants.

The authors benchmarked GPC against existing generative models and report improved performance on standard population genetics criteria including fidelity to linkage disequilibrium patterns and rare variant distributions. The study was published in PLOS Genetics on 6 October 2026.