Creating artificial human genomes using generative neural networks

Abstract

Generative models have shown breakthroughs in a wide spectrum of domains due to recent advancements in machine learning algorithms and increased computational power. Despite these impressive achievements, the ability of generative models to create realistic synthetic data is still under-exploited in genetics and absent from population genetics. Yet a known limitation in the field is the reduced access to many genetic databases due to concerns about violations of individual privacy, although they would provide a rich resource for data mining and integration towards advancing genetic studies. In this study, we demonstrated that deep generative adversarial networks (GANs) and restricted Boltzmann machines (RBMs) can be trained to learn the complex distributions of real genomic datasets and generate novel high-quality artificial genomes (AGs) with none to little privacy loss. We show that our generated AGs replicate characteristics of the source dataset such as allele frequencies, linkage disequilibrium, pairwise haplotype distances and population structure. Moreover, they can also inherit complex features such as signals of selection. To illustrate the promising outcomes of our method, we showed that imputation quality for low frequency alleles can be improved by data augmentation to reference panels with AGs and that the RBM latent space provides a relevant encoding of the data, hence allowing further exploration of the reference dataset and features for solving supervised tasks. Generative models and AGs have the potential to become valuable assets in genetic studies by providing a rich yet compact representation of existing genomes and high-quality, easy-access and anonymous alternatives for private databases.

Automatic inference of demographic parameters using generative adversarial networks

“…The use of GANs in population genetics is just beginning. Recently, Yelmen et al ( 2021 ) created a GAN that generates artificial genomes that mirror the properties of real genomes. Their approach does not include an evolutionary model, so the resulting artificial genomes are ‘unlabelled’.…”

Deep‐learning power and perspectives for genomic selection

“…For instance, when playing the Atari game, the computer or agent playing this game was positively rewarded when the outcome of the game was positive based on the actions performed. The algorithm was able to learn some of the games to a level where it performed better than humans (Patterson & Gibson, 2017 (Nielsen & Voigt, 2018 ); sequences (Linder et al, 2019;Liu et al, 2020;Yelmen et al, 2021 ); singlecell RNA sequencing data (Grønbech et al, 2020;Liu et al, 2020;Marouf et al, 2020 ); protein sequences (Repecka et al, 2021;Sinai et al, 2017 ); promoter sequences (Y, ; high-resolution Hi-C data (Hong et al, 2020;Liu et al, 2019b;Liu et al, 2020 ); among others. Also, VAE in GS has been applied to visualize population structure (Battey et al, 2021).…”

“…These methods are very efficient at creating fake images (or text) that, for humans, are identical to the real ones. In biology, these methods are being used to generate artificial genomes; fake DNA, such as microbial genomes (Nielsen & Voigt, 2018 ); sequences (Linder et al, 2019;Liu et al, 2020;Yelmen et al, 2021 ); singlecell RNA sequencing data (Grønbech et al, 2020;Liu et al, 2020;Marouf et al, 2020 ); protein sequences (Repecka et al, 2021;Sinai et al, 2017 ); promoter sequences (Y, ; high-resolution Hi-C data (Hong et al, 2020;Liu et al, 2019b;Liu et al, 2020 ); among others. Also, VAE in GS has been applied to visualize population structure (Battey et al, 2021).…”

Privacy considerations for sharing genomics data

“…1) ). In this regard, researchers have utilized the concept of generative modeling, applying for example Restricted Boltzmann Machines (RBMs) (Hinton et al, 2006[ 43 ]) or Generative Adversarial Networks (GANs) (Goodfellow et al, 2014[ 35 ]; Yelmen et al, 2021[ 81 ]). Though not with the goal to create entire synthetic datasets, variational autoencoders (VAEs) (Kingma and Welling, 2013[ 46 ]) have been used to impute missing counts in single-cell expression data (Eraslan et al, 2019[ 25 ]; Qiu et al, 2020[ 61 ]) as well as in bulk-RNA sequencing and methylome data (Qiu et al, 2020[ 61 ]).…”