Biology subjects

Mas Montserrat, D.

Publications and source records attributed to Mas Montserrat, D..

8 recordsLinked to original sources

bioRxiv2023

PopGenAdapt: Semi-Supervised Domain Adaptation for Genotype-to-Phenotype Prediction in Underrepresented Populations

The lack of diversity in genomic datasets, currently skewed towards individuals of European ancestry, presents a challenge in developing inclusive biomedical models. The scarcity of such data is particularly evident in labeled datasets that include genomic data linked to electronic health records. To address this gap, this paper presents PopGenAdapt, a genotype-to-phenotype prediction model which adopts semi-supervised domain adaptation (SSDA) techniques originally proposed for computer vision. PopGenAdapt is designed to leverage the substantial labeled data available from individuals of European ancestry, as well as the limited labeled and the larger amount of unlabeled data from currently underrepresented populations. The method is evaluated in underrepresented populations from Nigeria, Sri Lanka, and Hawaii for the prediction of several disease outcomes. The results suggest a significant improvement in the performance of genotype-to-phenotype models for these populations over state-of-the-art supervised learning methods, setting SSDA as a promising strategy for creating more inclusive machine learning models in biomedical research. Our code is available at https://github.com/AI-sandbox/PopGenAdapt.

bioinformatics↗

bioRxiv2023

Deep Variational Autoencoders for Population Genetics

MotivationModern biobanks provide numerous high-resolution genomic sequences of diverse populations. These datasets enable a better understanding of genotype-phenotype interactions with genome-wide association studies (GWAS) and power a new personalized precision medicine with polygenic risk scores (PRS). In order to account for diverse and admixed populations, new algorithmic tools are needed in order to properly capture the genetic composition of populations. Here we explore deep learning techniques, namely variational autoencoders (VAEs), to process genomic data from a population perspective. We hope this work will encourage the adoption of deep neural networks in the population genetics community. ResultsIn this paper, we show the power of VAEs for a variety of tasks relating to the interpretation, classification, simulation, and compression of genomic data with several worldwide whole genome datasets from both humans and canids and evaluate the performance of the proposed applications with and without ancestry conditioning. The unsupervised setting of autoencoders allows for the detection and learning of granular population structure and inferring of informative latent factors. The learned latent spaces of VAEs are able to capture and represent differentiated Gaussian-like clusters of samples with similar genetic composition on a fine-scale from single nucleotide polymorphisms (SNPs), enabling applications in dimensionality reduction, data simulation, and imputation. These individual genotype sequences can then be decomposed into latent representations and reconstruction errors (residuals) which provide a sparse representation useful for lossless compression. We show that different population groups have differentiated compression ratios and classification accuracies. Additionally, we analyze the entropy of the SNP data, its effect on compression across populations, its relation to historical migrations, and we show how to introduce autoencoders into existing compression pipelines.

bioinformatics↗

bioRxiv2022

Adversarial Attacks on Genotype Sequences

Adversarial attacks can drastically change the output of a method by performing a small change on its input. While they can be a useful framework to analyze worst-case robustness, they can also be used by malicious agents to perform damage in machine learning-based applications. The proliferation of platforms that allow users to share their DNA sequences and phenotype information to enable association studies has led to an increase in large databases. Such open platforms are, however, vulnerable to malicious users uploading corrupted genetic sequence files that could damage downstream studies. Such studies commonly include steps involving the analysis of the genomic sequences structure using dimensionality reduction techniques and ancestry inference methods. In this paper we show how white-box gradient-based adversarial attacks can be used to corrupt the output of genomic analyses, and we explore different machine learning techniques to detect such manipulations.

bioinformatics↗

bioRxiv2022

Predicting Dog Phenotypes from Genotypes

We analyze dog genotypes (i.e., positions of dog DNA sequences that often vary between different dogs) in order to predict the corresponding phenotypes (i.e., unique observed characteristics). More specifically, given chromosome data from a dog, we aim to predict the breed, height, and weight. We explore a variety of linear and non-linear classification and regression techniques to accomplish these three tasks. We also investigate the use of a neural network (both in linear and non-linear modes) for breed classification and compare the performance to traditional statistical methods. We show that linear methods generally outperform or match the performance of non-linear methods for breed classification. However, we show that the reverse is true for height and weight regression. Finally, we evaluate the results of all of these methods based on the number of input features used in the analysis. We conduct experiments using different fractions of the full genomic sequences, resulting in input sequences ranging from 20 SNPs to [~]200k SNPs. In doing so, we explore the impact of using a very limited number of SNPs for prediction. Our experiments demonstrate that these phenotypes in dogs can be predicted with as few as 0.5% of randomly selected SNPs (i.e., 992 SNPs) and that dog breeds can be classified with 50% balanced accuracy with as few as 0.02% SNPs (i.e., 40 SNPs).

genomics↗

bioRxiv2022

Generative Moment Matching Networks for Genotype Simulation

The generation of synthetic genomic sequences using neural networks has potential to ameliorate privacy and data sharing concerns and to mitigate potential bias within datasets due to under-representation of some population groups. However, there is not a consensus on which architectures, training procedures, and evaluation metrics should be used when simulating single nucleotide polymorphism (SNP) sequences with neural networks. In this paper, we explore the use of Generative Moment Matching Networks (GMMNs) for SNP simulation, we present some architectural and procedural changes to properly train the networks, and we introduce an evaluation scheme to qualitatively and quantitatively assess the quality of the simulated sequences.

bioinformatics↗

bioRxiv2022

Hyperbolic Geometry-Based Deep Learning Methods to Produce Population Trees from Genotype Data

The production of population-level trees using the genomic data of individuals is a fundamental task in the field of population genetics. Typically, these trees are produced using methods like hierarchical clustering, neighbor joining, or maximum likelihood. However, such methods are non-parametric: they require all data to be present at the time of tree formation, and the addition of new data points necessitates the regeneration of the entire tree, a potentially expensive process. They also do not easily integrate with larger workflows. In this study, we aim to address these problems by introducing parametric deep learning methods for tree formation from genotype data. Our models specifically create continuous representations of population trees in hyperbolic space, which has previously proven highly effective in embedding hierarchically structured data. We present two different architectures - a multi-layer perceptron (MLP) and a variational autoencoder (VAE) - and we analyze their performance using a variety of metrics along with comparisons to established tree-building methods. Both models tested produce embedding spaces that reflect human evolutionary history. In addition, we demonstrate the generalizability of these models by verifying that addition of new samples to an existing tree occurs in a semantically meaningful manner. Finally, we use Dasguptas Cost to compare the quality of trees generated by our models to those produced by established methods. Despite the fact that the benchmark methods are directly fit on the evaluation data, our models are able to outperform some of these and achieve highly comparable performance overall. Author summaryTree production is a vital task in population genetics, but current approaches fall prey to several common shortfalls. Most notably, they lack the ability to add new data points after tree generation, and they are often difficult to use in larger pipelines. By leveraging cutting-edge advances pairing deep learning with hyperbolic geometry, we develop multiple models designed to rectify these issues. Through experiments on a dataset of humans from globally widespread ancestries, we demonstrate the generalizability of our models to new data, and we also show strong empirical performance with respect to currently used methods. In addition, we show that the data representations produced by our models are semantically meaningful and reflect known facts about human evolutionary history. Finally, we discuss the additional benefits our models could provide, including improved visualization, greater privacy preservation, and improved integration with downstream machine learning tasks. In conclusion, we present models that are accurate, flexible, and generalizable, with the potential to facilitate a variety of further applications.

evolutionary biology↗

bioRxiv2021

Archetypal Analysis for Population Genetics

The estimation of genetic clusters using genomic data has application from genome-wide association studies (GWAS) to demographic history to polygenic risk scores (PRS) and is expected to play an important role in the analyses of increasingly diverse, large-scale cohorts. However, existing methods are computationally-intensive, prohibitively so in the case of nationwide biobanks. Here we explore Archetypal Analysis as an efficient, unsupervised approach for identifying genetic clusters and for associating individuals with them. Such unsupervised approaches help avoid conflating socially constructed ethnic labels with genetic clusters by eliminating the need for exogenous training labels. We show that Archetypal Analysis yields similar cluster structure to existing unsupervised methods such as ADMIXTURE and provides interpretative advantages. More importantly, we show that since Archetypal Analysis can be used with lower-dimensional representations of genetic data, significant reductions in computational time and memory requirements are possible. When Archetypal Analysis is run in this fashion, it takes several orders of magnitude less compute time than the current standard, ADMIXTURE. Finally, we demonstrate uses ranging across datasets from humans to canids. Author summaryThis work introduces a method that combines the singular value decomposition (SVD) with Archetypal Analysis to perform fast and accurate genetic clustering by first reducing the dimensionality of the space of genomic sequences. Each sequence is described as a convex combination (admixture) of archetypes (cluster representatives) in the reduced dimensional space. We compare this interpretable approach to the widely used genetic clustering algorithm, ADMIXTURE, and show that, without significant degradation in performance, Archetypal Analysis outperforms, offering shorter run times and representational advantages. We include theoretical, qualitative, and quantitative comparisons between both methods.

bioinformatics↗

bioRxiv2021

High Resolution Ancestry Deconvolution for Next Generation Genomic Data

As genome-wide association studies and genetic risk prediction models are extended to globally diverse and admixed cohorts, ancestry deconvolution has become an increasingly important tool. Also known as local ancestry inference (LAI), this technique identifies the ancestry of each region of an individuals genome, thus permitting downstream analyses to account for genetic effects that vary between ancestries. Since existing LAI methods were developed before the rise of massive, whole genome biobanks, they are computationally burdened by these large next generation datasets. Current LAI algorithms also fail to harness the potential of whole genome sequences, falling well short of the accuracy that such high variant densities can enable. Here we introduce Gnomix, a set of algorithms that address each of these points, achieving higher accuracy and swifter computational performance than any existing LAI method, while also enabling portable models that are particularly useful when training data are not shareable due to privacy or other restrictions. We demonstrate Gnomix (and its swift phase correction counterpart Gnofix) on worldwide whole-genome data from both humans and canids and utilize its high resolution accuracy to identify the location of ancient New World haplotypes in the Xoloitzcuintle, dating back over 100 generations. Code is available at https://github.com/AI-sandbox/gnomix.

genomics↗