Population labels can be generated directly from targeted next-generation sequencing data

doi:10.21203/rs.3.rs-5282595/v1

Download PDF

Brief Communication

Population labels can be generated directly from targeted next-generation sequencing data

https://doi.org/10.21203/rs.3.rs-5282595/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

The validity of genetic studies is reliant on the selection of appropriately matched population controls to prevent erroneous associations between population-specific genetic variants and disease. Such studies have traditionally relied on self-declared ethnicity which is likely to produce inaccurate predictions and is ethically problematic. More recently, ancestry informative markers (AIMs) have been used to determine the genetic similarity of an individual to ancestry reference populations. These AIMS, however, mostly reside in the non-coding DNA, making it difficult to determine ancestry from sequencing data which does not cover the whole genome. To address this, we implemented an empirical methodology that utilizes Procrustes analysis and a random forest classification to select genetically similar gnomAD control populations for study samples. This approach avoids the problems associated with using ethnicity as a substitute for genetic similarity and can be used to select suitable controls for studies that rely on exome or targeted sequencing data.

Biological sciences/Genetics/Population genetics

Biological sciences/Biotechnology/Sequencing

Using appropriately matched genetic controls is essential for accurate and reproducible genetic studies to mitigate the risk of invalid results due to population stratification issues¹. For example, when performing burden tests to explore variants’ enrichment in a disease cohort, the use of poorly matched controls can lead to detection of population differences in allele frequency rather than a disease association. Identifying appropriate controls can, however, be challenging.

Genetic similarity to ancestry reference populations can be determined using ancestry informative markers (AIMs), which are mostly located in non-coding genomic regions². These sites are comprised of common polymorphisms with different frequency distributions across different ancestral populations. In genetic ancestry analysis, dimensionality reduction techniques are used to turn genotype data at AIMs into a small number of continuous variables describing the genetic ancestry spectrum³. These variables are used to classify similarity to reference populations using machine learning models or techniques such as distance from the centre point of population clusters^4,5. However, in studies analysing data generated by targeted next-generation-sequencing (tNGS) or whole-exome-sequencing (WES), coverage of AIMs is often insufficient to use standard models to predict genetic ancestry. One means of improving coverage that has rarely been implemented and that we have explored in this study is the inclusion of off-target reads in the analysis. These are spread throughout the genome and exist as a byproduct of targeted capture, stemming from inefficiencies in the capture process.

Some studies have used self-declared ethnicity as a surrogate for genetic background; however, this practice has many limitations⁶. Ethnicity is a complex and subjective concept that is often conflated with other aspects of identity and is distinct from genetic ancestry. This is exemplified by an internationally referred subset of our cohort of individuals with monogenic disorders of insulin secretion, for whom over 300 distinct ethnicity terms, encompassing various concepts, such as race, religion, and geography, were provided on referral forms (Fig. 1). Self-reported ethnicity thus represents a poor approximation for genetic background.

The use of self-reported ethnicity is also ethically problematic: many terms have colonial origins⁷ and were subsequently perpetuated by the field of eugenics⁸. The recent publication of the National Academies report⁹ on population descriptors in science has been pivotal to addressing concerns related to the use of different descriptors. Thus, for practical and ethical reasons it is important to be able to identify the appropriate reference population for an individual based on their genetic data and not self-reported ethnicity.

Here, we describe the implementation of an empirical methodology for selecting genetically similar gnomAD controls¹⁰ for individuals sequenced by tNGS or WES without relying on information such as self-reported ethnicity.

Population Cohorts

We built a reference ancestry principal component (PC) space using the combined 1000 genomes and human genome diversity project (1000G + HGDP) whole-genome-sequencing (WGS) dataset¹¹ that forms part of the gnomAD database¹⁰ (N = 3,901 with gnomAD ancestry labels). This dataset was chosen as it is the only subset of the gnomAD database where individual-level data is available.

In the development and testing of the ancestry pipeline, we used a cohort of 7,509 individuals with a monogenic disorder of insulin secretion who had undergone tNGS analysis at the Exeter genomics laboratory^12–14 (Exeter-MDIS cohort). These individuals have been referred from 113 countries, representing a wide range of genetic backgrounds. WGS data was available for 381 of these individuals.

Development of the Classification Method

We used Plink 1.9¹⁵ to filter the combined 1000G + HGDP reference dataset to 848,202 AIMs based on frequency (minor allele frequency > 0.05), linkage disequilibrium (window size 100bp, step size 5, LD threshold 0.5) and missingness (missing genotype rate < 0.01). The number of AIMs here is higher than would typically be used in WGS-based ancestry analysis but maximises the possible number of sites that can be covered by random off-target reads. From this dataset, we calculated the first 10 PCs. Next, we used the LASER¹⁶ tool to perform Procrustes analysis and place the 7,509 individuals in the Exeter-MDIS cohort into our reference ancestry space. Procrustes analysis is a form of statistical shape analysis used here to identify the optimal translation, rotation, and scaling factors to translate a PC space created using just the AIMs covered by on- and off-target tNGS reads into the original reference space built using all 848,202 AIMs.

Finally, we created a random forest model for classification using the reference PC data and the gnomAD-provided ancestry labels for individuals in the population. To enhance the model's classification ability, we incorporated 10 rounds of self-training.¹⁷ In each round, the model iteratively classified 500 individuals from an ancestrally diverse subset of the Exeter-MDIS cohort selected based on kernel density across the PC space, incorporating those with a classification confidence of > 0.9 into the training set for the next round. This step aimed to help the model better understand the boundaries between different population groups.

Method Assessment

We used a correlation analysis to evaluate the effectiveness of the Procrustes step. We compared the PC values generated using the LASER Procrustes method on tNGS data to those generated using standard Plink 1.9¹⁵ PC projection on WGS data from 381 individuals for whom tNGS and WGS data was available.

To assess the accuracy of the model, we used it to classify a subset of 976 individuals in the gnomAD reference dataset who had not been included in the model training stage. The subset was selected randomly from the original reference population in a population stratified manner. We compared the classification output for the testing subset with the original population labels provided by gnomAD.

As an additional test of the model’s performance on unseen data, we performed population classification and UMAP clustering on the remaining 7,009 individuals in the Exeter-MDIS cohort who were not included in the model training. This was to ensure that the classifications were separated into distinct clusters and to check that the model had not overfitted to the training data.

tNGS derived PCs were concordant with those derived using standard WGS methods

We found a high correlation between PC values produced using tNGS data with the Procrustes method and those produced through standard PC projection in the WGS (Fig. 2). PCs 1–4 and 7 were among the highest correlated (Pearson’s R > 0.99, p < 2.2e^− 16) and were revealed as the most important for population classification by the random forest analysis.

The Random Forest ancestry classification model performed well on unseen PC data

The ancestry model had 99.8% accuracy in identifying the matching gnomAD-provided population label when tested on 976 individuals in the reference dataset that were not used in the creation of the model.

When we applied the pipeline to the 7,009 individuals from the Exeter-MDIS cohort not used in the model creation, 6,496 individuals (92.68%) were classified as genetically similar to a group within the control individuals. The classified populations were divided into clear, distinct clusters (Fig. 3).

We describe an empirical method for selecting gnomAD control groups with a similar genetic background to study samples sequenced using tNGS or WES. Pipeline testing using subsets of the Exeter-MDIS and 1000G + HGDP reference cohorts revealed it to be reliable and accurate. We show that by using this method, the most genetically similar population control group in gnomAD can be identified for a given sample of interest without resorting to suboptimal control selection methods like self-reported ethnicity.

In our evaluation of the LASER¹⁶ Procrustes method for projecting tNGS samples into a reference ancestry space, we found a strong correlation with the standard WGS-based PC projection methodology. This indicates that tNGS data contains sufficient on and off-target information to accurately plot the data points onto the 10 PCs generated for the WGS-derived reference and that the Procrustes method is proficiently projecting our study samples into the reference ancestry space.

The self-trained random forest classification model was tested with unseen 1000G + HGDP reference data and accurately reproduced the gnomAD reference labels. The model also performed well on unseen data from the Exeter-MDIS cohort, with clear separation into distinct clusters. These results are indicative of the model's ability to perform well on new and unseen data, without overfitting to the training data.

The pipeline was unable to classify 513 (7.3%) individuals, due to them not having a probability of > 0.75 for any single population. It is possible that these individuals have mixed ancestry or come from a region outside of those used to delineate the different gnomAD population groups. The pipeline output includes individual match probabilities for each gnomAD population group, enabling researchers to independently set their own threshold and decide how to treat individuals who match multiple populations.

The pipeline's inability to classify certain individuals may also be due to poor representation of their genetic ancestry in the reference data. Despite efforts by the International Genome Sample Resource to maximize diversity in the 1000G + HGDP reference cohort, some regions remain underrepresented¹⁸.

We describe a robust, reproducible method that enables the classification of samples with gnomAD population labels in studies reliant on tNGS or WES data. This approach removes the need for using unsuitable surrogates such as self-reported ethnicity.

Ethical Approval

The study complied with the Declaration of Helsinki, and the families of the children gave their informed written consent for genetic testing and recruitment to the Genetic Βeta-cell Research Bank (Exeter, UK; ethical approval was provided by the North Wales Research Ethics Committee, UK; IRAS project ID 231760).

This research was funded in whole, or in part, by Wellcome [223187/Z/21/Z]. For the purpose of open access, the author has applied a CC BY public copyright licence to any Author accepted Manuscript version arising from this submission.

Author contributions

J.R.S., T.W.L., M.N.W., M.B.J., S.E.F and E.D.F. participated in study conception and design. M.B.J., S.E.F, A.T.H., and E.D.F. recruited the cohort used in the model design and testing. J.R.S. developed the classification model. J.R.S. and M.H.B. carried out statistical analysis of the cohort data and model efficacy. J.R.S. wrote the first draft of the manuscript. T.W.L., M.N.W., M.B.J., S.E.F, E.D.F. and A.T.H. participated in manuscript improvement. All authors reviewed the manuscript.

Acknowledgements

This research was funded in whole, or in part, by the Wellcome Trust [223187/Z/21/Z and 224600/Z/21/Z]. For the purpose of open access, the author has applied a CC BY public copyright licence to any Author accepted Manuscript version arising from this submission. This research was supported by the National Institute for Health and Care Research (NIHR) Exeter Biomedical Research Centre (BRC) and National Institute for Health and Care Research Exeter Clinical Research Facility. The views expressed are those of the author(s) and not necessarily those of the NIHR or the Department of Health and Social Care. M.H.B. was funded by the Natural Environment Research Council (NERC). ATH is employed as a core member of staff within the National Institute for Health Research–funded Exeter Clinical Research Facility and is an NIHR Emeritus Senior Investigator. M.B.J. is funded by a Diabetes UK/Breakthrough T1D RD Lawrence Fellowship. S.E.F. has a Wellcome Trust Senior Research Fellowship [223187/Z/21/Z]. EDF is funded by a Diabetes UK RD Lawrence Fellowship [19/005971].

Cardon LR, Palmer LJ. Population stratification and spurious allelic association. Lancet 2003; 361: 598–604.
Enoch MA, Shen PH, Xu K, Hodgkinson C, Goldman D. Using ancestry-informative markers to define populations and detect population stratification. http://dx.doi.org/101177/1359786806066041 2006; 20: 19–26.
Rosenberg NA et al. Genetic structure of human populations. Science 2002; 298: 2381–2385.
Zhang W, Cheng L, Huang G. Towards fine-scale population stratification modeling based on kernel principal component analysis and random forest. Genes Genomics 2021; 43: 1143–1155.
Byun J, Han Y, Gorlov IP, Busam JA, Seldin MF, Amos CI. Ancestry inference using principal component analysis and spatial analysis: A distance-based analysis to account for population substructure. BMC Genomics 2017; 18: 1–12.
Ma IWY, Khan NA, Kang A, Zalunardo N, Palepu A. Systematic review identified suboptimal reporting and use of race/ethnicity in general medical journals. J Clin Epidemiol 2007; 60: 572–578.
Heinz A, Müller DJ, Krach S, Cabanis M, Kluge UP. The uncanny return of the race concept. Front Hum Neurosci 2014; 8: 1–10.
Yudell M, Roberts D, DeSalle R, Tishkoff S. Science and society: Taking race out of human genetics. Science (1979) 2016; 351: 564–565.
National Academies of Sciences E and M. Using Population Descriptors in Genetics and Genomics Research: A New Framework for an Evolving Field. Using Population Descriptors in Genetics and Genomics Research: A New Framework for an Evolving Field 2023; : 1–217.
Chen S et al. A genome-wide mutational constraint map quantified from variation in 76,156 human genomes. bioRxiv 2022; : 2022.03.20.485034.
Koenig Z et al. A harmonized public resource of deeply sequenced diverse human genomes. bioRxiv 2023. doi:10.1101/2023.01.23.525248.
Ellard S et al. Improved genetic testing for monogenic diabetes using targeted next-generation sequencing. Diabetologia 2013; 56: 1958–1963.
Yau D et al. Using referral rates for genetic testing to determine the incidence of a rare disease: The minimal incidence of congenital hyperinsulinism in the UK is 1 in 28,389. PLoS One 2020; 15. doi:10.1371/JOURNAL.PONE.0228417.
Pang L et al. Improvements in Awareness and Testing Have Led to a Threefold Increase Over 10 Years in the Identification of Monogenic Diabetes in the U.K. Diabetes Care 2022; 45: 642.
Chang CC, Chow CC, Tellier LCAM, Vattikuti S, Purcell SM, Lee JJ. Second-generation PLINK: rising to the challenge of larger and richer datasets. Gigascience 2015; 4: 7.
Wang C, Zhan X, Liang L, Abecasis GR, Lin X. Improved Ancestry Estimation for both Genotyping and Sequencing Data using Projection Procrustes Analysis and Genotype Imputation. Am J Hum Genet 2015; 96: 926–937.
Tanha J, van Someren M, Afsarmanesh H. Semi-supervised self-training for decision tree classifiers. International Journal of Machine Learning and Cybernetics 2017; 8: 355–370.
Mauleekoonphairoj J et al. A diverse ancestrally-matched reference panel increases genotype imputation accuracy in a underrepresented population. Scientific Reports 2023 13:1 2023; 13: 1–8.
Yu N et al. Larger Genetic Differences Within Africans Than Between Africans and Eurasians. Genetics 2002; 161: 269–274.

There is no duality of interest

supplementaryfigure1.ai
Supplementary Figure 1: Flow chart summarising the steps involved in creating the classification pipeline for identifying the most genetically similar ancestry control group in gnomAD for an individual sequenced using targeted next generation sequencing or whole exome sequencing.

Download PDF

Review #1 received at journal
18 Nov, 2024
Reviewer #2 agreed at journal
09 Nov, 2024
Reviewer #1 agreed at journal
23 Oct, 2024
Reviewers invited by journal
23 Oct, 2024
Submission checks completed at journal
17 Oct, 2024
First submitted to journal
17 Oct, 2024
Unknown event
17 Oct, 2024
Editor assigned by journal
17 Oct, 2024

You are reading this latest preprint version

Population labels can be generated directly from targeted next-generation sequencing data

Status:

Version 1

Abstract

Figures

Introduction

Methods

Population Cohorts

Development of the Classification Method

Method Assessment

Results

tNGS derived PCs were concordant with those derived using standard WGS methods

The Random Forest ancestry classification model performed well on unseen PC data

Discussion

Declarations

Ethical Approval

Author contributions

Acknowledgements

References

Additional Declarations

Supplementary Files

Status:

Version 1