bioRxiv2019
Genome Wide Association Studies (GWASs) have the potential to reveal the genetics of microbial phenotypes such as antibiotic resistance and virulence. Capitalizing on the growing wealth of bacterial sequence data, microbial GWAS methods aim to identify causal genetic variants while ignoring spurious associations. Bacteria reproduce clonally, leading to strong population structure and genome-wide linkage, making it challenging to separate true \"hits\" (i.e. mutations that cause a phenotype) from non-causal linked mutations. GWAS methods attempt to correct for population structure in different ways, but their performance has not yet been systematically evaluated. Here we developed a bacterial GWAS simulator (BacGWASim) to generate bacterial genomes with varying rates of mutation, recombination, and other evolutionary parameters, along with a subset of causal mutations underlying a phenotype of interest. We assessed the performance (recall and precision) of three widely-used univariate GWAS approaches (cluster-based, dimensionality-reduction, and linear mixed models, implemented in PLINK, pySEER, and GEMMA) and one relatively new whole-genome elastic net model implemented in pySEER, across a range of simulated sample sizes, recombination rates, and causal mutation effect sizes. As expected, all methods performed better with larger sample sizes and effect sizes. The performance of clustering and dimensionality reduction approaches to correct for population structure were considerably variable according to the choice of parameters. Notably, the elastic net whole-genome model was consistently amongst the highest-performing methods and had the highest power in detecting causal variants with both low and high effect sizes. Most methods reached good performance (Recall > 0.75) to identify causal mutations of strong effect size (log Odds Ratio >= 2) with a sample size of 2000 genomes. However, only elastic nets reached reasonable performance (Recall = 0.35) for detecting markers with weaker effects (log OR [~]1) in smaller samples. Elastic nets also showed superior precision and recall in controlling for genome-wide linkage, relative to univariate models. However, all methods performed relatively poorly on highly clonal (low-recombining) genomes, suggesting room for improvement in method development. These findings show the potential for whole-genome models to improve bacterial GWAS performance. BacGWASim code and simulated data are publicly available to enable further comparisons and benchmarking of new methods.\n\nAuthor summaryMicrobial populations contain measurable phenotypic differences with important clinical and environmental consequences, such as antibiotic resistance, virulence, host preference and transmissibility. A major challenge is to discover the genes and mutations in bacterial genomes that control these phenotypes. Bacterial Genome-Wide Association Studies (GWASs) are family of methods to statistically associate phenotypes with genotypes, such as point mutations and other variants across the genome. However, compared to sexual organisms such as humans, bacteria reproduce clonally meaning that causal mutations tend to be strongly linked to other mutations on the same chromosome. This genome-wide linkage makes it challenging to statistically separate causal mutations from non-causal false-positive associations. Several GWAS methods are currently available, but it is not clear which is the most powerful and accurate for bacteria. To systematically evaluate these methods, we developed BacGWASim, a computational pipeline to simulate the evolution of bacterial genomes and phenotypes. Using simulated genomes, we found that GWAS methods varied widely in their performance. In general, causal mutations of strong effect (e.g. those under strong selection for antibiotic resistance) could be easily identified with relatively small samples sizes of around 1000 genomes, but more complex phenotypes controlled by mutations of weaker effect required 3000 genomes or more. We found that a recently-developed GWAS method called elastic net was particularly good at identifying causal mutations in highly clonal populations, with strong linkage between mutations - but there is still room for improvement. The BacGWASim computer code is publicly available to enable further comparisons and benchmarking of new methods.