1. Epidemiology and Global Health
  2. Genetics and Genomics
Download icon

Polygenic Scores: How well can we separate genetics from the environment?

  1. Jennifer Blanc
  2. Jeremy J Berg  Is a corresponding author
  1. Human Genetics, University of Chicago, United States
  • Cited 0
  • Views 2,447
  • Annotations
Cite this article as: eLife 2020;9:e64948 doi: 10.7554/eLife.64948


A simulation study demonstrates a better method for separating genetic effects from environmental effects in genome-wide association studies, but there is still some way to go before this becomes a "solved" problem.

Main text

A person’s traits – such as their height or risk of disease – result from a complex interplay between the genes they inherit and the environments they experience over their lifetime. To cut through some of this complexity, human geneticists use a tool called a polygenic score, which attempts to predict a person’s traits solely from their genes (Rosenberg et al., 2019).

To build a polygenic score, geneticists first enroll a large number of people in a genome-wide association study (GWAS). For each participant, researchers measure numerous genetic variants across their genome, together with a trait of interest, and use this data to determine the extent to which different variants are associated with the trait. This information makes it possible to take the genome of someone who was not involved in the original GWAS and add up the effects of multiple genetic variants to calculate a polygenic score for that trait (Figure 1A). These scores have been used to predict a person’s risk of developing a disease (Torkamani et al., 2018), to study our evolutionary past (Rosenberg et al., 2019), and to help understand complex social outcomes (Harden and Koellinger, 2020).

Correcting biases in polygenic scores.

(A) A genome-wide associate study (GWAS) measures the trait of interest (phenotype) and the genotype of a sample of individuals and uses this data (middle graph) to see which genetic variants (represented by individual dots) are associated with the trait of interest (shown in red). This information is used to compute the polygenic score of individuals not in the original sample. Individuals with a higher polygenic score (orange) are predicted to have a higher trait value (e.g. to be taller or to have a greater risk of disease), while those with a lower polygenic score are predicted to have a lower trait value (bottom graph). (B) Mathieson and Zaidi simulated genetic data for a population that separated into subpopulations in the recent past; the environment was simulated as a six-by-six grid (left) in which environmental factors associated with the trait of interest vary smoothly from top to bottom. The uncorrected mean polygenic scores (top right) have a structure that clearly mirrors the structure in the environment. Correcting the scores with the 'common PCA' approach (middle right) does not solve this problem, but correction with the 'rare PCA' approach (bottom right) does. (C) However, when differences in the environmental factors were localized to a single square in the grid (shown in yellow), not even the rare PCA model could eliminate the correlation between genetic and environmental effects (indicated by asterix).

Image credit: Panel A – top (Stux, CC0), middle (Figure 1, Hu et al., 2016, CC BY 4.0), bottom (Jennifer Blanc); Panel B (Adapted from Figure 4, Zaidi and Mathieson, 2020).

However, efforts to use polygenic scores face substantial obstacles. All human populations exhibit genetic structure – variation in how genetically similar pairs of individuals are to one another – due to the complex history of geographic separation, population mixtures and migrations that have occurred throughout our evolutionary history. If this genetic structure correlates with patterns of environmental variation, it will cause many genetic variants to be incorrectly associated with a trait. This phenomenon, which is known as population stratification, will introduce biases into polygenic scores and undermine their purpose (which is to separate out the genetic component of trait variation).

To overcome this barrier, researchers would ideally measure the relevant environmental effects in the GWAS sample and include them as statistical controls in their analyses. However, it is difficult – if not impossible – to quantify all environmental effects on a given trait. Existing theory suggests that researchers can use the patterns of genetic variation they have already measured to model the genetic structure of the GWAS sample, and use this as statistical control instead (Song et al., 2015; Wang and Blei, 2019). In essence, because the problem arises from correlations between the environmental effects and patterns of genetic structure, it can be solved by controlling for either of them. The difficulty lies in how to correctly model this genetic structure. Geneticists favor a method called principal components analysis (PCA) (Price et al., 2006), as its simplicity and computational feasibility make it easy to apply to massive GWAS datasets. But the approach has limitations, and population stratification remains an issue in practice (Mathieson and McVean, 2012; Berg et al., 2019; Sohail et al., 2019).

Now, in eLife, Arslan Zaidi and Iain Mathieson from the University of Pennsylvania report which PCA models are the most effective at reducing bias in polygenic scores (Zaidi and Mathieson, 2020). To do this, they simulated the genetic data of a single population which had divided into spatially structured sub-groups within the recent past. They then simulated environmental effects on the trait and tested different PCA models to see how well each model controlled for them.

The results showed that the usual approach, known as ‘common PCA’, leads to polygenic scores that inappropriately mirror the environmental effects. Common PCA models calculate genetic structure by only measuring variants that appear in more than 5% of individuals in the GWAS sample. These common variants are typically ancient in origin, and therefore do not adequately capture the genetic structure of populations which have been spatially subdivided in the recent past. It is this failure to capture the genetic structure that results in biased polygenic scores.

On the other hand, rare variants, which appear in only a handful of individuals, are typically recent in origin and therefore reflect the history of recent subdivisions. Zaidi and Mathieson show that for this reason, PCA models built using patterns of genetic structure in rare variants (‘rare PCA’) eliminate biases from polygenic scores more effectively than the ‘common PCA’ technique (Figure 1B). However, this approach is not a panacea. When the environmental factors associated with the trait were localized to one geographic place (e.g. pollution localized to a particular city), even the rare PCA approach could not separate genetic effects from environmental biases (Figure 1C).

Zaidi and Mathieson also explore a more complicated set of simulations which are meant to more accurately mimic the patterns seen in real GWAS datasets, and find that the results are essentially identical to the simplified scenario described above. In all of their simulations, Zaidi and Mathieson know the ground truth, allowing them to experiment with different approaches designed to target the kind of bias they have simulated. In the real world, the ground truth is not known, so it is difficult to have complete confidence that stratification biases have been properly dealt with. Although a long-studied issue, these findings further demonstrate how separating genetic effects from environmental effects is still not a ‘solved’ problem in genetic studies (Lawson et al., 2020).

Studies that use polygenic scores have exploded in number over the past decade, riding a wave of well-founded optimism that they can open up new, otherwise inaccessible, avenues of research. But care is needed to ensure that this powerful tool is applied appropriately. Ultimately, the possibility for misleading results is an unavoidable risk, especially in research that is restricted to non-experimental settings. Zaidi and Mathieson provide several good recommendations for overcoming this, and suggest that a combination of the rare and common PCA approaches will minimize the amount by which environmental effects confound GWAS data. Moving forward, their results highlight the need for further statistical methods that more effectively deal with the biases introduced by environmental effects, especially for sharply distributed factors. In addition, more sensitive diagnostics are needed to assess how environmental effects impact polygenic scores.


  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
    The blessings of multiple causes
    1. Y Wang
    2. DM Blei
    Journal of the American Statistical Association 114:1574–1596.
  12. 12

Article and author information

Author details

  1. Jennifer Blanc

    Jennifer Blanc is in the Department of Human Genetics, University of Chicago, Chicago, United States

    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-7569-018X
  2. Jeremy J Berg

    Jeremy J Berg is in the Department of Human Genetics, University of Chicago, Chicago, United States

    For correspondence
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5411-6840


We thank Arjun Biddanda, Xiaoheng Cheng, Graham Coop, Doc Edge and John Novembre for comments on earlier drafts, and Arslan Zaidi and Iain Mathieson for answering questions about their paper.

Publication history

  1. Version of Record published: December 23, 2020 (version 1)


© 2020, Blanc and Berg

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.


  • 2,447
    Page views
  • 172
  • 0

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Epidemiology and Global Health
    Ting-Ting Zhao et al.
    Research Article

    Opisthorchiasis is an overlooked danger to Southeast Asia. High-resolution disease risk maps are critical but haven't been available for Southeast Asia. Georeferenced disease data and potential influencing factor data were collected through a systematic review of literatures and open-access databases, respectively. Bayesian spatial-temporal joint models were developed to analyze both point- and area-level disease data, within a logit regression in combination of potential influencing factors and spatial-temporal random effects. The model-based risk mapping identified areas of low, moderate and high prevalence across the study region. Even though the overall population-adjusted estimated prevalence presented a trend down, a total of 12.39 million (95% BCI: 10.10-15.06) people were estimated infected with O. viverrini in 2018 in four major endemic countries (i.e., Thailand, Laos, Cambodia, and Vietnam), highlighting the public health importance of the disease in the study region. The high-resolution risk maps provide valuable information for spatial targeting of opisthorchiasis control interventions.

    1. Epidemiology and Global Health
    2. Evolutionary Biology
    Jonathan Stieglitz et al.
    Research Article Updated

    In comparative cross-species perspective, humans experience unique physical impairments with potentially large consequences. Quantifying the burden of impairment in subsistence populations is critical for understanding selection pressures underlying strategies that minimize risk of production deficits. We examine among forager-horticulturalists whether compromised bone strength (indicated by fracture and lower bone mineral density, BMD) is associated with subsistence task cessation. We also estimate the magnitude of productivity losses associated with compromised bone strength. Fracture is associated with cessation of hunting, tree chopping, and walking long distances, but not tool manufacture. Age-specific productivity losses from hunting cessation associated with fracture and lower BMD are substantial: ~397 lost kcals/day, with expected future losses of up to 1.9 million kcals (22% of expected production). Productivity loss is thus substantial for high strength and endurance tasks. Determining the extent to which impairment obstructs productivity in contemporary subsistence populations improves our ability to infer past consequences of impairment.