Main

Over the past decade, genome-wide association studies (GWAS) have yielded remarkable advances in the understanding of complex traits and have identified hundreds of genetic risk variants in humans (for examples, see Refs 1, 2, 3). GWAS analyse hundreds of thousands to millions of common genetic variants, usually single-nucleotide polymorphisms (SNPs), to test for an association between each variant and a phenotype of interest (see Ref. 4). GWAS have confirmed the heritability of many human traits5, clarified their underlying genetic architecture6, and have identified novel biological mechanisms and drug targets7. Of recent interest to infectious disease researchers are microbial GWAS, which identify risk variants on the genomes of microorganisms, such as bacteria, viruses and protozoa. With increasingly cheap and high-throughput sequencing technologies, microorganism whole-genome sequences (WGS) are now being generated on an unprecedented scale that rivals human data. Microbial GWAS provide a new opportunity to develop insights into the biological mechanisms that underlie clinical outcomes, such as drug resistance and pathogenesis. As in human GWAS, insights from microbial GWAS may lead to the identification of molecular targets for drug and vaccine development. Furthermore, identifying genetic variants through microbial GWAS will enable researchers to track the evolution and spread of pathogenic strains through populations and to synthesize microorganisms in vitro that have the desired clinical phenotypes.

Human GWAS provide an optimistic outlook for microbial GWAS. However, there are important differences between microbial and human genomic studies that could hinder the success of microbial GWAS or require methodological adaptations. In this Review, we first outline specific features of GWAS methods and consider their application to microorganisms. Second, we summarize the microbial GWAS that have been carried out to date, outlining their key findings, methods and challenges. Although these studies have mainly focused on pathogenic viruses, bacteria and protozoa, and thus are the dominate focus of this Review, it is important to note that the same methods can also be applied to non-pathogenic microorganisms. Finally, we discuss the lessons that have been learned from human GWAS and anticipate the future of microbial GWAS, particularly the opportunities provided by the ability to collect GWAS data from both the host and microorganisms.

Data and methodology of GWAS

GWAS grew from the common disease common variant (CDCV) hypothesis8, which postulates that many high-frequency but low-effect variants contribute to disease risk. This hypothesis explained how diseases can avoid selection, manifest in complex inheritance patterns, and be genetically and phenotypically heterogeneous. GWAS aim to identify the common variants that underpin the heritability observed for many phenotypes9 (Box 1). These common variants are usually in the form of bi-allelic SNPs, where two nucleotides (A, C, G or T) exist at a locus with a frequency of more than 1% in the population. Each SNP is analysed, usually through linear or logistic regression, to determine whether one allele is significantly associated with the phenotype. Effects are reported as either beta for quantitative traits or odds ratio for case–control studies. Typically, only the main effects of individual SNPs are calculated, as methods for the detection of epistatic interactions between SNPs and SNP–environment interactions are challenging owing to the additional burden of multiple testing10,11. The power of the human GWAS approach came from genotyping chips that enable the rapid calling of hundreds of thousands of SNPs from across an individual's genome. Owing to the co-inheritance of segments of the genome over generations, correlations (known as linkage disequilibrium (LD)) exist between genetic variants that are in close proximity. LD allows genotyping chips to 'tag' local genetic variation by including a single proximal SNP, and to impute additional SNPs that were not directly genotyped based on known correlations12.

There are several differences between human GWAS and microbial GWAS (Table 1), one of the most important of which is the source of the genomic data. Unlike human GWAS, for which the data come from SNP genotyping chips, almost all genomic data for microorganisms come from sequencing. This affects several aspects of GWAS, particularly SNP calling, as SNPs that are detected in microbial sequencing data will not only be bi-allelic, but also tri-allelic and quad-allelic. This complicates variant calling, data storage and analysis. Matching loci to a reference genome is also of increased importance in microbial GWAS, to ensure that SNPs are called at the same location for each sample and for comparison across studies. Sequencing also affects the quality control steps that must be taken to filter SNPs and individual samples. Owing to the large number of SNPs compared with the number of samples in a study, quality control is carried out to preferentially exclude low-quality SNPs. Standard quality control in human GWAS removes the SNPs with low minor allele frequency (with a typical cut-off ranging from <1% to 5%), high missingness (>1–5%), and the SNPs that are out of Hardy–Weinberg equilibrium (P < E-5 or -6). Quality control on individual samples in a human study also removes samples with a high missingness (>1–5%) or that are outliers in genome-wide homozygosity. With the exception of Hardy–Weinberg equilibrium, these same quality control metrics will remain important for microbial GWAS. However, quality control thresholds need to be established for additional metrics that capture the quality of sequencing, such as sequencing depth and Phred scores.

Table 1 Conceptual and analytical steps of human GWAS and microbial GWAS

Adapting GWAS to microbial variants

As mentioned above, human GWAS typically focus on the effects of individual SNPs. However, focusing on the effects of SNPs alone will not always be possible in microbial GWAS. For example, in bacteria, recombination can introduce novel genes. This means that the causative genetic difference may be the presence or absence of an entire gene or set of genes. Microbial GWAS need to test this variation in gene presence alongside SNPs. In this case, lessons may come from the analysis of copy number variants (CNVs) in human GWAS. CNVs are large duplications or deletions of sections of the genome. CNV analyses test for associations between a phenotype and both specific CNVs and — owing to the rarity of specific CNVs — an individual's CNV burden. An individual's CNV burden is the proportion of their entire genome, or a region of it, that is covered by CNVs13. Similarly, analyses of human sequence data often test for associations with the burden of rare variants14. The contribution of variants to that burden can be weighted by their predicted functional impact. Using quantitative burdens that combine the effects of multiple genetic variants into a single variable might provide statistical methods for analysing gene presence or absence and rare variants in microbial GWAS.

Another approach to handling gene presence in microbial GWAS is defining and analysing k-mers15. The benefit of k-mers is that they simultaneously capture common variation and gene presence. Analysis of k-mers may also be useful owing to the larger proportion of coding sequence that is found in many microorganisms compared with humans, where only a small proportion of DNA is exonic. This is because k-mers can capture multiple allele differences that code for different amino acids, and thus reflect changes closer to the biological mechanism that underlies the phenotype of interest.

It is worth noting that most human GWAS have focused on the additive effects of variants. This is where each additional copy of an allele carried by a diploid organism increases the likelihood of a phenotype in a linear manner. However, owing to within-host evolution and the possibility of superinfection, some microorganisms will exhibit within-host genetic diversity. Within-host diversity will lead to non-discrete SNP calling, where the frequency of an allele reflects its frequency on microbial sequences within the host, rather than the presence or absence of an allele. Although testing for a linear association between allele frequency and phenotype makes pragmatic sense, the possibility of nonlinear effects also exists. Further, within-host diversity results in alleles from different lineages having unique LD patterns within the same host. This will be relevant to the analysis of epistatic interactions, as alleles within the same host may have different genetic backgrounds.

Finally, microbial GWAS are also likely to observe lineage effects. In this case, entire lineages, such as viral subtypes, might differ in phenotype. Thus, the lineage or subtype of the microorganism might be the genetic unit of interest, either alone or in addition to the effects of individual SNPs or k-mers. Disentangling the effects of a single variant from the effects that are related to lineage is potentially challenging, but has been shown to increase the power of microbial GWAS when mplemented successfully16.

Confounding factors in microbial GWAS

The main challenge that is associated with GWAS is the risk of identifying seemingly causal variants that are in fact false positives17. This is due to two main causes: population structure and multiple testing (see below). The use of samples from within a genetically diverse population can lead to subtle confounding from population structure, for example, because of an excess of cases from one ethnic group. In such instances, GWAS would identify predictive SNPs that are only informative of ancestry, rather than the biology of the disease. To avoid this problem, human GWAS often restrict recruitment to ethnically homogeneous groups. Even within relatively homogeneous populations, some population structure will exist. These subtler influences of population stratification are corrected through principal component analysis. This generates covariates that capture SNP correlations across the genome, and can be carried out using software such as EIGENSTRAT18. Principal components can capture subtle ancestry differences with high accuracy and can identify samples that represent population outliers19. Although principal components will be key to removing confounding that is due to population structure in microbial GWAS, two additional confounders exist that may require additional methods.

The first of these is homologous recombination, which occurs in bacteria and viruses through the replacement of short sequence blocks, rather than through multiple crossovers along the whole chromosome. This means that long-range LD is broken down differently in microbial genomes, leaving variants in long-range LD with each other even when short-range LD within a region is reduced20. This long-range LD could make the identification of the causal variant problematic21. Methods that are designed for analysing historically ethnically mixed, or 'admixed', human populations may be helpful in this case, because they make use of recombination patterns to identify associated loci22.

The second source of confounding is that microbial population structure can represent selection on the phenotype of interest, for example, antibiotic resistance. Given the differences in frequency of recombination and selection across microorganisms, the consequent population structures are likely to range from purely clonal to nearly panmictic. In addition, the rapid spread of successful lineages may temporarily reduce their recombination with the rest of the species. In microorganisms in which there has been strong selection, it may be appropriate to use repeated samples from within a single host over time, such as comparing pretreatment and post-treatment sequences. However, this approach will not work for longitudinal phenotypes, such as the time taken to develop disease symptoms, or in microorganisms with low rates of evolution. In these studies, methods that use mixed models to account for relatedness15 or lineage effects16, or to identify signals of selection across the genome based on phylogenetic structure23, may have more traction than typical GWAS regression methods.

Multiple testing and replication

Aside from confounding, the other major source of false positives is the multiple testing that is intrinsic to GWAS. The standard cut-off for an association to be considered statistically significant is P = 0.05, which represents a 5% probability of random occurrence. However, testing hundreds of thousands of SNPs leads to tens of thousands of SNPs being significant at P < 0.05 by chance alone. To account for the number of tests, a SNP must pass the genome-wide significance cut-off in order to be considered significant (Box 2). This is usually P < 5E-8 in humans24, which is approximately equal to the Bonferroni correction (a multiple testing correction) for the number of SNPs analysed in early GWAS. However, it continues to be used in more densely genotyped and imputed studies. Additional SNPs included in GWAS through deeper genotyping or imputation are in high LD with those already known, and so the correlations between SNPs reduces the number of independent tests carried out. Thus, understanding the level of LD between SNPs is important for calculating the correct threshold for genome-wide significance. Even with strict cut-offs for genome-wide significance, determining whether an association represents a false positive remains problematic.

As a result, replication in an independent cohort is the gold standard for reporting an association in GWAS25. This is both to avoid false positives and to accurately estimate the effect size of the SNP. Normally, GWAS have reduced power to detect variants of small effect and there is consequently a bias towards identifying novel SNPs that have an over-estimated effect size (sometimes called the 'winner's curse')26. As no bias for discovery exists during replication, the effect size in the replication cohort will more accurately reflect the true effect. Generally, replication does not require the association of a SNP to reach genome-wide significance in the replication cohort, but to pass a P value threshold based on the number of SNPs brought forward for replication. Further, meta-analysis of the P values of a SNP in both the discovery and the replication cohorts should surpass genome-wide significance in order for a SNP to be considered a true positive.

However, microbial GWAS may be less reliant on replication than human GWAS given that suspected causal variants can be validated in vitro. This ability to generate carriers of identified variants and to test their effect in the laboratory reduces many of the concerns of false positives that are typically associated with human GWAS. It also provides model organisms that can be used to gain a better understanding of the function of the variant. One important area of research is the development of methods to identify and correct for epistasis. Epistasis can take the form of specific interactions between two SNPs or the effect of a SNP being conditional on a broader genetic background. Disentangling epistatic effects will be key to generating viable in vitro models of microbial GWAS findings and establishing causality.

Power, polygenicity and heritability

As well as providing methodological insights, the history of GWAS predicts a clear trajectory for how progress in microbial GWAS is likely to unfold. Initial human GWAS identified only a small number of SNPs, each explaining only a tiny fraction of variation. The disparity between expected heritability from twin studies and the heritability explained by genome-wide significant associations became known as the 'missing heritability' (Ref. 27). Missing heritability initially cast doubt on the GWAS approach. However, as the first waves of studies were pooled into meta-analyses28, and the second waves of GWAS were analysed, more and more associations were reported, increasing the amount of heritability explained29. It became clear that the stringent cut-off for statistical significance resulted in a need for larger sample sizes than had been expected in order to achieve sufficient power to identify SNPs. Once sufficient power was reached, the relationship between the sample size and number of SNPs identified became relatively linear. However, despite this, there was often an inverse relationship between the frequency of identified SNPs and their effect size, meaning that each SNP explained only a small fraction of variation29.

The problem of missing heritability persisted, leading to a move away from single SNP analyses and towards polygenic methods30 (Fig. 1). One of the first polygenic methods was the use of polygenic risk scores (PRSs)31. PRSs are based on the assumption that many SNPs with small effect sizes will fail the stringent cut-off that is used for genome-wide significance; however, together their cumulative effect could explain a large amount of the variance in risk. The construction of a PRS requires both a discovery and a replication cohort. In the discovery cohort, a GWAS is carried out, defining the 'risk' allele and effect size of each SNP regardless of whether the P value is significant. In the replication cohort, the number of 'risk' alleles that an individual sample carries is summed into a score (the PRS), with each allele weighted by its effect size. The variation in case–control status that is predicted by the PRS is then calculated. Several PRSs are often defined using different P value thresholds for the inclusion of SNPs from the discovery GWAS, for example, four scores using SNPs with P < 0.001, P < 0.05, P < 0.2 and P < 0.5. As more SNPs are included, there is a greater likelihood that all SNPs of true effect will be included. However, including more SNPs also increases the number of SNPs with no true effect, and thus adds noise, which causes the amount of variance that is explained to plateau. PRSs ultimately provide a more powerful predictive tool than the results of single SNPs. As such, PRSs may be key to rapidly translating the results from microbial GWAS to prediction in the clinic, even before the roles of individual risk variants are understood.

Figure 1: Phenotype prediction as GWAS sample sizes increase.
figure 1

Variance in a phenotype (schizophrenia3) explained in successive waves of genome-wide association studies (GWAS) by the genome-wide significant (GW-Sig) single-nucleotide polymorphisms (SNPs) and polygenic risk scores (PRSs) from all SNPs with P < 0.05. As can be seen, the number of SNPs identified exponentially increases with sample size, and at every stage PRSs provide substantially better prediction than the use of significant SNPs alone. However, the challenge of 'missing heritability' continues even within fairly large GWAS, with the variance explained still below the heritability estimates derived from GREML and twin studies. The number of cases shown reflects the discovery sample size for the PRS analysis carried out.

PowerPoint slide

An alternative polygenic method is genomic-relatedness-matrix residual maximum likelihood analysis (GREML), which was often referred to in the early literature by the software name GCTA5. GREML estimates the proportion of variance that is captured by all SNPs and calculates the heritability of the phenotype. This is done by calculating how genetically similar each possible combination of two samples is (that is, their genetic relatedness). Relatedness refers to how much of the genome is shared between two samples (that is, they have the same genotypes). The heritability is then calculated as the proportion of phenotypic similarity between samples that can be explained by their relatedness. It is important to note that GREML does not estimate the true heritability of a phenotype, it estimates only the heritability that is captured by the included SNPs. Unlike PRS, GREML does not provide a means of predicting risk. However, it does act as a benchmark for the maximum amount of risk that is detectable in an infinitely powered GWAS. For example, in humans, GREML was used to estimate that common SNPs account for between one-third and one-half of the heritability estimated from twin studies30 (Fig. 1). Although PRS and GREML have not been widely used in microorganisms, they will be key to understanding whether current microbial GWAS are underpowered and whether novel variants will be identified with larger sample sizes.

A crucial aspect of polygenic methods is their ability to identify what drives the heritability of a phenotype. First, polygenic methods can be used to test whether heritability is disproportionately driven by specific genomic regions, by rare or common variants, or by variants within particular biological pathways. Second, polygenic methods can measure the heritability of specific subtypes of the phenotype. Identifying phenotypic subtypes with higher heritability identifies individuals for whom the microbial genome is most relevant. Furthermore, polygenic methods are able to identify a genetic correlation between two phenotypes, even when data are available on only one phenotype in each sample32. Thus, they can determine whether two distinct phenotypes have overlapping aetiologies, or whether two subtypes of a phenotype are genetically distinct. Polygenic analyses have supported the generalist genes hypothesis, according to which genetic effects are highly pleiotropic33. Overall, human GWAS predict that, for traits under moderate selection, the genetic architecture will consist of many small effect and pleiotropic variants, which are spread fairly evenly across allele frequencies and genomic regions.

Progress in microbial GWAS

Given the clear trajectory of human GWAS from underpowered studies to more advanced methods that explain a significant proportion of risk, it makes sense to ask whether microbial GWAS will advance in the same manner. Despite the complexities mentioned above, a growing number of microbial GWAS have recently been published (Table 2). With the exception of HIV and Plasmodium falciparum, these publications have generally focused on bacteria and have almost exclusively focused on pathogens within human hosts. Most genomic data have come from WGS, although genotyping chips for P. falciparum have existed for several years34,35. Owing to the much shorter genomes of microorganisms, the number of variants analysed in microbial GWAS has been in the tens of thousands, which is orders of magnitude smaller than in human GWAS. Sample sizes have also been considerably smaller. The smallest microbial GWAS so far was a study of 75 Staphylococcus aureus strains36 and the largest was a study of 3,701 Streptococcus pneumoniae isolates37. The majority of studies have had sample sizes of less than 500 (Table 2). However, this promises to change as large multi-country consortia, such as MalariaGEN38 and PANGEA_HIV39, generate WGS on a much larger scale.

Table 2 Examples of microbial GWAS

Despite the current small sample sizes, microbial GWAS have already been successful in identifying causal variants. This is partly due to the studies focusing on phenotypes that are under strong selection, the majority of which were studies on drug resistance. For example, microbial GWAS of Mycobacterium tuberculosis40, S. aureus36, S. pneumoniae37, P. falciparum41 and HIV have all successfully identified novel drug resistance variants that often explained almost all of the phenotypic variation. Even with phenotypes under strong selection, there has been evidence of high polygenicity within microorganisms. For example, the study of drug resistance in 3,701 S. pneumoniae sequences identified 301 significant SNPs, with a median odds ratio of 11 (Ref. 37). Given the large effect sizes, it is not surprising that many of the drug resistance variants that were identified through microbial GWAS were previously known. Although this diminishes the novelty of the findings, it also strengthens confidence in the ability of microbial GWAS to correctly identify causal variants. Another phenotype under strong selection is host specificity. Microbial GWAS of host specificity have yielded significant results for Campylobacter jejuni42 and HIV43. However, within the same study of HIV host specificity, the authors found no associations between viral variants and infectiousness. The most successful study of virulence was of 90 S. aureus samples44. The authors identified 121 SNPs at genome-wide significance. Functional follow-up of a subset of SNPs showed that four of 13 affected toxicity in vivo, suggesting that a proportion of the associations identified were truly causal.

Most microbial GWAS have so far focused on the analysis of traits that are under strong selection, but these studies have shown remarkable diversity in their analytical approaches (Fig. 2). Two analyses of HIV sequences have been carried out43,45, both using the GWAS software PLINK46. On the basis of fixed-effect models, these studies suggested that the virus shows low levels of population stratification within a single viral subtype. However, analyses of M. tuberculosis highlighted that although PLINK could identify many drug resistance variants, it also led to false positives owing to confounding from population structure47.To address this limitation, the authors developed the software PhyC23, a tool that uses phylogenetic trees to identify SNPs under recent convergent evolution. This approach identified many of the same drug resistance variants as PLINK, but reduced the level of confounding from population structure. Other studies have included phylogenetic structure as a random effect in mixed models, using software such as ROADTRIPS48 and FaST-LMM49. These mixed models have successfully reduced the effect of population structure in a number of microorganisms36,41. One of the limitations of this software is that these programs are designed for human genomic data and cannot handle features such as within-host microbial diversity. A recent study developed a bespoke approach to microbial GWAS in the analysis of C. jejuni42. The authors generated multi-allelic k-mers, rather than SNPs, and tested these for an association with host preference. This is the only study so far to combine an analysis of SNPs with gene presence or absence, which is a key genomic feature of bacteria.

Figure 2: Potential models for microbial GWAS.
figure 2

Examples of three microbial genome-wide association study (GWAS) approaches to date40,41,43. a | The organism analysed in each study: HIV, a retrovirus that causes AIDS; Plasmodium falciparum, a parasitic protozoa that is the cause of malaria; and Mycobacterium tuberculosis, a bacterium that causes tuberculosis. b | The form of geographic, population or phylogenetic confounding observed in each organism, which hinders the ability to differentiate single-nucleotide polymorphisms (SNPs) of true effect from systematic false positives. For HIV, only minimal population structure was observed, whereas for P. falciparum greater population differences existed. M. tuberculosis showed the highest level of confounding, with the different phenotypes (represented by the red and white nodes of the phylogenetic tree) mostly clustering within the same lineages. c | Given the different population and phylogenetic structures of the three organisms, three different approaches were used to carry out the microbial GWAS. The lack of confounding in HIV allowed for the application of typical human GWAS fixed-effect models. The more substantial population structure in P. falciparum was accounted for by including phylogenetic relatedness as a random effect in a mixed model. Finally, the clear phylogenetic structure of M. tuberculosis was used to carry out genome-wide analysis of convergent selection. d | How the results of each microbial GWAS were taken forwards to better understand the microorganism. For HIV, the viral genomic data were combined with human GWAS data to carry out a genome-to-genome analysis of HIV viral load. For P. falciparum, the information on drug resistance variants was combined with geographic data to highlight the spread of resistance variants through Southeast Asia. Finally, for M. tuberculosis, the identified drug resistance variant (Δald) was functionally validated by showing that carriers had improved growth comparable to other resistant strains (Bacillus Calmette–Guérin (BCG)) and sensitivity was partially resotored by complementation (Δald-comp), to levels similar to those of the wild type (WT). BD, Bangladesh; MM, Myanmar; TH, Thailand; LA, Laos; VN, Vietnam. The left part of panel d is adapted from Ref. 43. The middle part of panel d is from Ref. 41, Nature Publishing Group. The right part of panel d is from Ref. 40, Nature Publishing Group.

PowerPoint slide

Overall, it is clear that although microbial GWAS are yielding important insights into infectious disease, the field has yet to settle on a consistent analytical approach and current methods are not yet ideally suited to microbial genomes. More refined analytical methods will become particularly important as the focus of microbial GWAS expands beyond drug resistance and towards phenotypes in which variants have subtler polygenic effects.

Remaining lessons

As microbial GWAS become more widespread, there are still several lessons that can be learned from human GWAS. Perhaps the most crucial lesson revolves around the generation of sufficient sample sizes to identify variants of small effect. This requires a collaborative approach. Samples must often be pooled from across the world in order to create well-powered discovery and replication cohorts. Of particular note is the mega-analytic approach that pools raw genotype data from all sites into a central repository, which is used for standardized quality control and to increase power50. There are good reasons for optimism as international microbial research consortia already exist.

One area that has not yet been explored in microbial GWAS is the trade-off between sample size and heterogeneity. As more complex phenotypes are analysed, heterogeneity will reduce power to detect the causal variants. With finite resources and time, there is a choice between focusing on collecting detailed clinical data on a smaller number of more homogeneous samples, and recruiting large numbers of samples with minimal screening. In human GWAS, both approaches have been shown to be effective. First, power can be improved by restricting to 'super controls' (Ref. 51), for example, using controls on the opposite extreme of the phenotype of interest, or focusing on a subset of samples with a phenotype that is believed to be more homogeneous or heritable52,53. Second, 'minimal phenotyping' can be used to maximize sample size, such as assuming all those with records of treatment are ill54. Widely collected proxy phenotypes, such as education level as a proxy for cognitive ability, have been successfully used to maximize sample sizes for more complex traits55. Aetiologically similar phenotypes can also be jointly analysed to maximize sample size2,56. Overall, a sensible first step seems to be to increase sample sizes as much as possible. This can then be followed by secondary analyses of more homogeneous phenotypic subtypes in cases for which data are available.

Finally, many advances in human GWAS were made possible by free and open software applications (such as GCTA5 and PLINK46) that could handle various data formats and could carry out multiple analyses (Table 3). These software applications were generally very user friendly, with detailed documentation. Microbial GWAS have so far been carried out using a range of software with different analytical approaches (Table 3). Although GWAS software that can handle large genomic data sets already exists, these programs are not ideally suited to the non-diploid multi-allelic nature of some microbial genomes, and cannot carry out longitudinal within-individual sequence comparisons that might be desired. In particular, GWAS methods will need to be adapted to deal with within-host microbial diversity and recombination. Further, the successful polygenic methods for estimating the heritability and co-heritability of phenotypes from GWAS data have yet to be evaluated in microbial GWAS. As can be seen from GCTA5, a single piece of software with a topical application has driven a large number of high-profile advances in human genomics. The development of free and open software applications that can accurately and conveniently analyse a wide range of microbial WGS data to detect single SNP and polygenic effects is, therefore, a top priority of the field.

Table 3 Features of software applications used in microbial GWAS to date

Future directions: integrating the host

Arguably, the most exciting application of microbial GWAS is to integrate it with human genomic data. Human GWAS of infectious disease have been carried out for more than 12 pathogens (reviewed in Ref. 57). This Review ends by highlighting the potential for combining these findings with those of microbial GWAS. These genome-to-genome analyses can provide important insights into whether the effects of microbial variants are universal or whether they are dependent on a specific host genetic background. Such statistical host–microbial interactions would help to identify which host proteins the microorganism is interacting with on a molecular level. Further, interactions that prevent infection or disease progression would represent potential drug or vaccine targets.

The authors are aware of only one comprehensive genome-to-genome analysis at this time. The microbial GWAS of HIV set point viral load, mentioned above, generated both HIV sequences and host GWAS data43. This study was able to identify many associations between viral genetic variants and those in the human genome, specifically within the major histocompatibility complex region. In a secondary analysis, the importance of host–pathogen correlations and how they might lead to overestimates of the combined host and pathogen heritabilities were highlighted58. In this case, although both host and viral heritability of HIV set point viral load were observed, the two were shown to substantially overlap.

With cheaper genome-sequencing methods, the ability of groups to generate both host and microbial data on the same individuals will only increase. However, just as microbial GWAS currently lack universal analytic software, so do genome-to-genome analyses. Such statistical tools will be needed in order for the field to flourish, particularly as the scale of data will make these analyses computationally intensive. A simpler method may be to condense multiple SNPs into a single variable, as seen in PRS31, and to test for interactions on a genome-wide level. Regardless of the method used, the availability of host and microorganism GWAS data presents an opportunity to increase power to identify causal variants. Ideally, such data will be generated within large longitudinal studies, for which genomic data can also be combined with epidemiological and clinical variables. Understanding the correlations between host demography, host heritability and microorganism heritability will provide greater insights into the extent to which microbial genomes drive clinical outcomes.

Conclusions

As this Review has discussed, there is great promise in the field of microbial GWAS. However, it is clear that a number of analytical advances will be needed to handle the unique features of microbial genomics. Perhaps the issue of greatest importance will be the development of software applications that can handle the combined analysis of host and microorganism genomic data. With these tools, we will be better able to predict individual patient outcomes, track the evolution of global epidemics, and identify new drug and vaccine targets.