Admix-kit: An Integrated Toolkit and Pipeline for Genetic Analyses of Admixed Populations

Summary:

Admixed populations, with their unique and diverse genetic backgrounds, are often underrepresented in genetic studies. This oversight not only limits our understanding but also exacerbates existing health disparities. One major barrier has been the lack of efficient tools tailored for the special challenges of genetic study of admixed populations. Here, we present admix-kit, an integrated toolkit and pipeline for genetic analyses of admixed populations. Admix-kit implements a suite of methods to facilitate genotype and phenotype simulation, association testing, genetic architecture inference, and polygenic scoring in admixed populations.

Introduction

Admixed individuals inherit a mosaic of ancestry segments originating from multiple continental ancestral populations, leading to their complex and diverse genetic backgrounds encompassing a wide spectrum of human genetic variation (Seldin et al., 2011). An understanding of such genetic ancestry mosaicism within admixed populations offers opportunities to gain insights into the origins and health implications of various genetic traits and diseases, contributing to a more comprehensive understanding of human genetics (Wojcik et al., 2019; Tan and Atkinson, 2023).

Despite the genetic richness and crucial insights they can offer, admixed populations remain significantly underrepresented in current genetic studies (Mills and Rahal, 2020). This underrepresentation can be attributed to various challenges, including the complexity of analyzing diverse genetic backgrounds and the lack of efficient tools and standardized practices for handling the genetic data of admixed populations. This gap not only hinders progress in genetic research but also exacerbate health disparities. For example, findings with datasets from European ancestry groups for genetic risk prediction models can introduce bias to personalized risk prevention strategies (Martin et al., 2019; Ding et al., 2023).

To address these challenges, we introduce admix-kit, an integrated and flexible python toolkit along with workflows developed using Workflow Development Language (WDL), specifically designed for the simulation and analysis of genetic data from admixed populations. We anticipate that our proposed software packages and workflows will help overcome these analytical challenges, enabling the inclusion of admixed individuals in future genetic studies.

Results

Computational toolkit for analyzing admixed genotypes

We begin by outlining the data structures and computational tools in admix-kit for analyzing admixed genetic datasets. Both genotype and local ancestry data are organized as two matrices of shape N x M x 2 (N and M denote the number of individuals and SNPs respectively, and ‘2’ denotes the two haplotypes; Figure 1a). Given that storage of these matrices often exceeds memory capacity (large N and/or M), we adopt a chunked array representation, implemented with the Dask python library (Rocklin, 2015). Each chunk is loaded from disk on demand, thus conserving memory by loading data only when needed and facilitating large-scale analyses. We employ the pgenlib Python API (URLs) to read phased genotype. Local ancestry matrices are stored in a compressed format that leverages their contiguous nature (local ancestry for nearby SNPs are often identical within each individual). By translating genotype and local ancestry matrices into local-ancestry-specific (LAS) genotype dosages, we have also implemented a set of utility functions tailored for LAS genetic analysis, including LAS allele frequencies, polygenic scores and phenotype modeling that allow for LAS genetic architecture (Figure 1b).

Figure 1: Overview of admix-kit’s data structure, functionality, and illustrative analyses using a simulated dataset.

(a) Local ancestry and phased genotypes are stored in matrix format. Individual-specific and SNP-specific covariates are stored as two tables with matching orders. (b) Analysis based on ancestry-specific genotype dosage. Starting with a phased genotype for an individual (0/1 denotes presence of minor allele, blue/red color denotes European/African local ancestry), genotypes are separated into ancestry-specific dosages. Local ancestry-informed downstream analyses can be subsequently performed. (c) visualization of local ancestry tracts. (d) Consistency of genome-wide genetic ancestry of simulated dataset (e) Consistency of allele frequencies from the simulated admixed genotypes.

Workflow for simulating admixture genotypes

Genotype simulation is essential to facilitate testing and benchmarking genetic analysis methodologies. One of the significant challenges lies in simulating admixed genomes, which often becomes the most time-consuming step among common analyses involving admixture. We develop a workflow to specifically address this bottleneck (Figure S1a). We primarily focus on two-way admixture for demonstration while noting our software and pipeline are adaptable to various admixture scenarios. First, we use HAPGEN2 (Su et al., 2011) to expand sets of unique haplotypes within each reference genetic ancestry group (e.g. European/African), preserving the minor allele frequency (MAF) and linkage disequilibrium (LD) structure. Second, using the expanded haplotype sets in both genetic ancestry groups, we simulate the admixture process employing admix-simu (URLs) with parameters for genetic ancestry proportion and the number of admixture generations. This process mimics random mating and recombination events to generate realistic distribution of local ancestry segments, MAF and LD structure for the generated genotypes. To make this simulation process more accessible, we have implemented these functionalities as command-line tools within admix-kit (Figure S1a). In details, admix hapgen2 --pfile ${src_plink2} --n-indiv ${n_indiv} --out ${expanded_pop} is used to expand the source population with HAPGEN2. And admix admix-simu --pfile-list “[‘pop1’, ‘pop2’]” --admix-prop “[0.2,0.8]” --n-indiv ${n_indiv} --n-gen ${n_gen} --out ${admix} can be used to simulate the admixture process across source populations. Furthermore, a number of functions are implemented to enable LAS analysis including association testing (Pasaniuc et al., 2011; Atkinson et al., 2021; Hou et al., 2021; Mester et al., 2023) (admix assoc), genetic architecture inference (Hou et al., 2023) (admix genet-cor) and polygenic scoring (Marnetto et al., 2020; Sun et al., 2022) (admix calc-partial-pgs). We also implemented a userfriendly WDL-based workflow for genotype simulation that can be run on cloud-based computing platforms (e.g., AnVIL [https://anvilproject.org/], BioData Catalyst [https://biodatacatalyst.nhlbi.nih.gov/]) (Figure S1b). Users can input essential parameters to define the admixture process and provide the input genotype path of ancestral populations (a set of preprocessed 1,000 genomes dataset is provided for default usage). The workflow will run through each aforementioned step and produce the simulated admixed genotype dataset. The admix-kit software is encapsulated in a publicly available docker image (URLs).

Example analysis of a simulated dataset

We demonstrate the practicality of admix-kit through analyses of a simulated dataset. All associated code and notebooks have been made publicly accessible (https://github.com/UWGAC/admix-kit_workflow). This ensures our results are fully reproducible and can be seamlessly deployed in a cloud platform (e.g., AnVIL). We used the AnVIL workflow to simulate N = 1,000 admixed individuals with M = 174K SNPs on chromosome 1 and 2, using a demographic model similar to African American individuals with over 8 generations of admixture and an average ancestry proportion of 80% African and 20% European (Kidd et al., 2012) (ancestry proportion varies by individual). Notably, the genotype simulation took less than 30 minutes with scalability to a much larger number of individuals and SNPs. Using principal component analysis (PCA), we observed that individuals within the simulated dataset are positioned along a cline between individuals labeled as European and African in the 1,000 Genomes reference dataset, suggesting high quality of the simulated genotype dataset (Figure 1c–d). Allele frequencies computed within genotype segments corresponding to the respective local ancestry displayed high consistency with those computed in the reference population, indicating high preservation of MAF structure of the simulated genotype (Figure 1e).

Discussion

Addressing the underrepresentation of admixed individuals in genetic studies is pivotal not only for scientific necessity but also as a commitment to equity. With this goal in mind, we introduce admix-kit, a comprehensive toolkit and workflow tailored for admixed populations. We anticipate that our software package and workflows will facilitate greater inclusion of admixed individuals in future genetic studies.

Development of software and methodology in genetic studies relies heavily on the use of simulated datasets. These datasets help benchmark performance and facilitate comparisons with existing software. Traditionally, simulated datasets are usually derived from publicly available reference populations. Often, these populations are selected based on a high degree of genetic similarity among individuals in the population (e.g., individuals having all four grandparents from a small geographic region.) For instance, HAPGEN2 has recently been widely used for simulating large-scale genetic datasets that mimic the LD structures of reference populations such as European, African, American, East Asian, and South Asian using data from the 1000 Genomes Project (Su et al., 2011; Ruan et al., 2022; Zhang et al., 2022; Miao et al., 2023). While these simulations can recreate datasets with homogeneous LD as the reference populations, they may not capture the complex population structures seen in admixed populations (Figure S2). Consequently, these sampling conditions are not representative of global human genetic variations. As a remedy, simulating admixture among reference populations can provide datasets that more rigorously test the performance of new software. For example, our simulation pipeline can be used to investigate factors that potentially impact accuracy of ancestry inference (including ancestry composition in reference panel, demographic model of simulated admixed population and error in inferred local ancestry) and to understand how errors in ancestry inference propagate to downstream disease mapping and prediction applications.

Admix-kit holds significant potentials in the development of Polygenic Risk Scores (PRS). The efficacy of PRS is known to hinge on the similarity of the target population to the training population (Ding et al., 2023). With the PRIMED consortium working on methods to improve the performance of PRS in diverse populations, simulations will be pivotal for method evaluation (Kachuri et al., 2023). In this context, we expect that admix-kit will be an essential part of this effort.

Polygenic Risk Methods in Diverse Populations (PRIMED) Consortium Methods Working Group

Sally Adebamowo¹, Adebowale Adeyemo², Paul Auer³, Taoufik Bensellak², Sonja Berndt², Rohan Bhukar⁴, Hongyuan Cao⁵, Clinton Cario⁶, Nilanjan Chatterjee⁷, Jiawen Chen⁸, Tinashe Chikowore⁹, Ananyo Choudhury⁹, Matthew Conomos¹⁰, David Conti¹¹, Sinead Cullina¹², Burcu Darst¹³, Yi Ding¹⁴, Ruocheng Dong¹⁵, Rui Duan¹⁶, Yasmina Fakim¹⁷, Nora Franceschini⁸, Tian Ge¹⁸, Anisah W. Ghoorah¹⁷, Chris Gignoux¹⁹, Stephanie Gogarten¹⁰, Neil Hanchard², Rachel Hanisch², Michael Hauser²⁰, Scott Hazelhurst⁹, Jibril Hirbo²¹, Whitney Hornsby¹⁸, Kangcheng Hou¹⁴, Xing Hua², Alicia Huerta²², Micah Hysong⁸, Jin Jin²³, Angad Johar²⁴, Jon Judd⁶, Linda Kachuri⁶, Abram Bunya Kamiza⁹, Eimear Kenny¹², Alyna Khan¹⁰, Elena Kharitonova⁸, Joohyun Kim²¹, Iain Konigsberg¹⁹, Charles Kooperberg¹³, Matt Kosel²⁴, Iftikhar Kullo²⁴, Ethan Lange¹⁹, Yun Li⁸, Qing Li², Maria Liivrand²⁵, Kirk Lohmueller¹⁴, Kevin Lu²¹, Ravi Mandla⁴, Alisa Manning⁴, Iman Martin², Alicia Martin⁴, Shannon McDonnell²⁴, Leah Mechanic², Josep Mercader⁴, Rachel Mester¹⁴, Maggie Ng²¹, Kevin Nguyen¹, Kristján Norland²⁴, Franklin Ockerman⁸, Loes Olde Loohuis¹⁴, Ebuka Onyenobi¹, Bogdan Pasaniuc¹⁴, Aniruddh Patel⁴, Ella Petter¹⁴, Kenneth Rice¹⁰, Joseph Rothstein¹², Bryce Rowan¹², Robb Rowley², Yunfeng Ruan⁴, Sriram Sankararaman¹⁴, Ambra Sartori⁷, Dan Schaid²⁴, Ruhollah Shemirani¹², Jonathan Shortt¹⁹, Xueling Sim²⁶, Johanna Smith²⁴, Maggie Stanislawski¹⁹, Daniel Stram¹¹, Quan Sun⁸, Bamidele Tayo²⁷, Buu Truong⁴, Kristin Tsuo⁴, Sarah Urbut¹⁸, Ying Wang⁴, Wallace Minxian Wang⁴, Riley Wilson², John Witte⁶, Genevieve Wojcik⁷, Jingning Zhang⁷, Ruyue Zhang⁸, Haoyu Zhang², Yuji Zhang¹, Michael Zhong¹, Laura Zhou⁸

¹University of Maryland Baltimore , Baltimore, MD, United States, ²National Institutes of Health, Bethesda, MD, United States, ³Medical College of Wisconsin, Milwaukee, WI, United States, ⁴Broad Institute, Cambridge, MA, United States, ⁵Florida State University, Tallahassee, FL, United States, ⁶Stanford University, Stanford, CA, United States, ⁷Johns Hopkins University, Baltimore, MD, United States, ⁸University of North Carolina at Chapel Hill, Chapel Hill, NC, United States, ⁹University of the Witwatersrand, Johannesburg South Africa, ¹⁰University of Washington, Seattle, WA, United States, ¹¹University of Southern California, Los Angeles, CA, United States, ¹²Mount Sinai, New York City, NY, United States, ¹³Fred Hutchinson Cancer Center, Seattle, WA, United States, ¹⁴University of California Los Angeles, Los Angeles, CA, United States, ¹⁵University of Wisconsin Milwaukee, Milwaukee, WI, United States, ¹⁶Harvard, Cambridge, MA, United States, ¹⁷University of Mauritius, Réduit Mauritius, ¹⁸Massachusetts General Hospital, Boston, MA, United States, ¹⁹University of Colorado, Aurora, CO, United States, ²⁰Duke University, Durham, NC, United States, ²¹Vanderbilt University Medical Center, Nashville, TN, United States, ²²Instituto Nacional de Ciencias Médicas y Nutrición Salvador Zubiran, Mexico City, CDMX, Mexico, ²³University of Pennsylvania, Philadelphia, PA, United States, ²⁴Mayo Clinic, Rochester, MN, United States, ²⁵Genevia Technologies, Tampere Finland, ²⁶National University of Singapore, Singapore, ²⁷Loyola University of Chicago, Chicago, IL, United States