Ruth Massey and colleagues report an analysis used to predict the toxicity of methicillin-resistant Staphylococcus aureus (MRSA) isolates based on a genetic signature (Genome Res. doi:10.1101/gr.165415.113, 9 April 2014). They characterized 90 independent MRSA ST239 isolates for adhesiveness and toxicity, which are critical virulence factors. They found limited interstrain variation in binding to fibrinogen and fibronectin, whereas there was an 18-fold difference between the most and least toxic strains. To examine the relationship between toxicity and disease severity in vivo, they infected mice with the isolates with the highest and lowest toxicity, demonstrating that the more toxic isolates caused the most severe disease. They conducted a genome-wide association analysis for toxicity, identifying 100 SNPs and 22 indels associated at P < 0.05. Next, the authors screened for evidence of repeated independent evolution and segregation in clusters of highly toxic isolates, narrowing the candidate loci to four. They also tested for epistatic interactions between toxicity-associated loci. Finally, they built a predictive model for toxicity classification that uses a random forest machine learning algorithm and is based on data for a set of 31 SNPs and 21 indels. Using this model, they were able to correctly predict all 27 strains classified as having either high or low toxicity in this study, with lower accuracy shown in predicting strains of medium toxicity.