Biology subjects

Charpiat, G.

Publications and source records attributed to Charpiat, G..

6 recordsLinked to original sources

bioRxiv2026

Attentive-SPIDNA: Attention-based neural networks for population genetics

Artificial neural networks (ANNs) have recently offered new perspectives to solve inference problems from high dimensional data in numerous scientific fields, but it is yet unclear which architectures are the most suited to genomic data. Here, we present a new ANN architecture integrating attention mechanisms to infer effective population size history from genomic data. Built upon our previous exchangeable architecture SPIDNA, Attentive-SPIDNA adds attention layers that allow computing more expressive and complex features from combinations of haplotypes. The contribution of each haplotype to the features is learned automatically and depends on its content and affinity with the other haplotypes. Likewise, we use this mechanism to automatically perform a voting scheme that aggregates predictions from different genomic regions. This new architecture outperforms approximate Bayesian computation and previously published neural networks while relying directly on raw genetic data and being invariant to haplotype permutation in the input. As a proof-of-concept, we use this architecture to infer the effective population size history of 54 populations from the HGDP dataset (Bergstrom et al, 2020). This application highlights the ability of the network to handle data with a varying number of haplotypes and to quickly perform predictions for datasets including numerous populations. Therefore, the proposed mechanism could be integrated to various neural networks solving population genetics tasks.

evolutionary biology↗

bioRxiv2025

Bias in genome-wide association test statistics due to omitted interactions

Over the past two decades, genome-wide association studies (GWAS) enabled the discovery of thousands of variants associated with many complex human traits. However, conventional GWAS are still widely performed with linear models with the assumption that the genetic effects are predominantly additive. In this work, we investigate the test statistic behavior when linear models are used to obtain significant genotype-phenotype associations without accounting for epistasis. We first algebraically derive mean and variance shift in the null statistic due to the omitted interaction term, and define the boundary between conservative (i.e., deflated statistic tail) and anti-conservative (i.e., inflated statistic tail) regimes for the common GWAS significance threshold. We then perform phenotype simulation analyses using the Estonian Biobank genotypes and validate the mathematical model. We demonstrate that the anti-conservative regime is plausible under realistic parameter settings and models omitting interaction terms can produce spurious significance. Our findings suggest caution when interpreting statistically significant signals reported in the literature based on linear models, especially for large-scale GWAS.

bioinformatics↗

bioRxiv2024

Diffusion-based artificial genomes and their usefulness for local ancestry inference

The creation of synthetic data through generative modeling has emerged as a significant area of research in genomics, offering versatile applications from tailoring functional sequences with specific attributes to generating high-quality, privacy-preserving in silico genomes. Notwithstanding these advancements, a key challenge remains: while some methods exist to evaluate artificially generated genomic data, comprehensive tools to assess its usefulness are still limited. To tackle this issue and present a promising use case, we test artificial genomes within the framework of population genetics and local ancestry inference (LAI). Building on previous work in deep generative modeling for genomics, we introduce a novel, frugal diffusion model and show that it produces high-quality genomic data. We then assess the performance of a downstream machine learning LAI model trained on composite datasets comprising both real and/or synthetic data. Our findings reveal that the LAI model achieves comparable performance when trained exclusively on real data versus high-quality synthetic data. Moreover, we highlight how data augmentation using high-quality artificial genomes significantly benefits the LAI model, particularly when real data is limited. Finally, we compare the conventional use of a single synthetic dataset to a robust ensemble approach, wherein multiple LAI models are trained on diverse synthetic datasets, and their predictions are aggregated. Our study highlights the potential of frugal diffusion-based generative models and synthetic data integration in genomics. This approach could improve fair representation across populations by overcoming data accessibility challenges, while ensuring the reliability of genomic analyses conducted on artificial data.

bioinformatics↗

bioRxiv2024

Latent generative modeling of long genetic sequences with GANs

Synthetic data generation via generative modeling has recently become a prominent research field in genomics, with applications ranging from functional sequence design to high-quality, privacy-preserving artificial in silico genomes. Following a body of work on Artificial Genomes (AGs) created via various generative models trained with raw genomic input, we propose a conceptually different approach to address the issues of scalability and complexity of genomic data generation in very high dimensions. Our method combines dimensionality reduction, achieved by Principal Component Analysis (PCA), and a Generative Adversarial Network (GAN) learning in this reduced space. Using this framework, we generated genomic proxy datasets for very diverse human populations around the world. We compared the quality of AGs generated by our approach with AGs generated by the established models and report improvements in capturing population structure, linkage disequilibrium, and metrics related to privacy leakage. Furthermore, we developed a frugal model with orders of magnitude fewer parameters and comparable performance to larger models. For quality assessment, we also implemented a new evaluation metric based on information theory to measure local haplotypic diversity, showing that generative models yield higher diversity than real genomes. In addition, we addressed the shrinkage issue associated with PCA and generative modeling, examined its relation to the nearest neighbor resemblance metric, and proposed a resolution. Finally, we evaluated the effect of different binarization methods on the quality of the output AGs.

genomics↗

bioRxiv2023

Deep convolutional and conditional neural networks for large-scale genomic data generation

Applications of generative models for genomic data have gained significant momentum in the past few years, with scopes ranging from data characterization to generation of genomic segments and functional sequences. In our previous study, we demonstrated that generative adversarial networks (GANs) and restricted Boltzmann machines (RBMs) can be used to create novel high-quality artificial genomes (AGs) which can preserve the complex characteristics of real genomes such as population structure, linkage disequilibrium and selection signals. However, a major drawback of these models is scalability, since the large feature space of genome-wide data increases computational complexity vastly. To address this issue, we implemented a novel convolutional Wasserstein GAN (WGAN) model along with a novel conditional RBM (CRBM) framework for generating AGs with high SNP number. These networks implicitly learn the varying landscape of haplotypic structure in order to capture complex correlation patterns along the genome and generate a wide diversity of plausible haplotypes. We performed comparative analyses to assess both the quality of these generated haplotypes and the amount of possible privacy leakage from the training data. As the importance of genetic privacy becomes more prevalent, the need for effective privacy protection measures for genomic data increases. We used generative neural networks to create large artificial genome segments which possess many characteristics of real genomes without substantial privacy leakage from the training dataset. In the near future with further improvements in haplotype quality and privacy preservation, large-scale artificial genome databases can be assembled to provide easily accessible surrogates of real databases, allowing researchers to conduct studies with diverse genomic data within a safe ethical framework in terms of donor privacy. Author summaryGenerative modelling has recently become a prominent research field in genomics, with applications ranging from functional sequence design to characterization of population structure. We previously used generative neural networks to create artificial genome segments which possess many characteristics of real genomes but these segments were short in size due to computational requirements. In this work, we present novel generative models for generating artificial genomes with larger sequence size. We test the generated artificial genomes with multiple summary statistics to assess the haplotype quality, overfitting and privacy leakage from the training dataset. Our findings suggest that although there is still room for improvement both in terms of genome quality and privacy preservation, convolutional architectures and conditional generation can be utilised for generating good quality, large-scale genomic data. In the near future with additional improvements, large-scale artificial genomes can be used for assembling surrogate biobanks as alternatives to real biobanks with access restrictions, increasing data accessibility to researchers around the globe.

genomics↗

bioRxiv2020

Deep learning for population size history inference: design, comparison and combination with approximate Bayesian computation

AO_SCPLOWBSTRACTC_SCPLOWFor the past decades, simulation-based likelihood-free inference methods have enabled researchers to address numerous population genetics problems. As the richness and amount of simulated and real genetic data keep increasing, the field has a strong opportunity to tackle tasks that current methods hardly solve. However, high data dimensionality forces most methods to summarize large genomic datasets into a relatively small number of handcrafted features (summary statistics). Here we propose an alternative to summary statistics, based on the automatic extraction of relevant information using deep learning techniques. Specifically, we design artificial neural networks (ANNs) that take as input single nucleotide polymorphic sites (SNPs) found in individuals sampled from a single population and infer the past effective population size history. First, we provide guidelines to construct artificial neural networks that comply with the intrinsic properties of SNP data such as invariance to permutation of haplotypes, long scale interactions between SNPs and variable genomic length. Thanks to a Bayesian hyperparameter optimization procedure, we evaluate the performance of multiple networks and compare them to well established methods like Approximate Bayesian Computation (ABC). Even without the expert knowledge of summary statistics, our approach compares fairly well to an ABC based on handcrafted features. Furthermore we show that combining deep learning and ABC can improve performance while taking advantage of both frameworks. Finally, we apply our approach to reconstruct the effective population size history of cattle breed populations.

evolutionary biology↗