Computational approaches in infectious disease research

Abstract

Due to overuse and misuse of antibiotics, the global threat of antibiotic resistance is a growing crisis. Three critical issues surrounding antibiotic resistance are the lack of rapid testing, treatment failure, and evolution of resistance. However, with new technology facilitating data collection and powerful statistical learning advances, our understanding of the bacterial stress response to antibiotics is rapidly expanding. With a recent influx of omics data, it has become possible to develop powerful computational methods that make the best use of growing systems-level datasets. In this work, I present several such approaches that address the three challenges around resistance. While this body of work was motivated by the antibiotic resistance crisis, the approaches presented here favor generalization, that is, applicability beyond just one context. First, I present ShinyOmics, a web-based application that allow visualization, sharing, exploration and comparison of systems-level data. An overview of transcriptomics data in the bacterial pathogen Streptococcus pneumoniae led to the hypothesis that stress-susceptible strains have more chaotic gene expression patterns than stress-resistant ones. This hypothesis was supported by data from multiple strains, species, antibiotics and non-antibiotic stress factors, leading to the development of a transcriptomic entropy based, general predictor for bacterial fitness. I show the potential utility of this predictor in predicting antibiotic susceptibility phenotype, and drug minimum inhibitory concentrations, which can be applied to bacterial isolates from patients in the near future. Predictors for antibiotic susceptibility are of great value when there is large phenotypic variability across isolates from the same species. Phenotypic variability is accompanied by genomic diversity harbored within a species. I address the genomic diversity by developing BFClust, a software package that for the first time enables pan-genome analysis with confidence scores. Using pan-genome level information, I then develop predictors of essential genes unique to certain strains and predictors for genes that acquire adaptive mutations under prolonged stress exposure. Genes that are essential offer attractive drug targets, and those that are essential only in certain strains would make great targets for very narrow-spectrum antibiotics, potentially leading the way to personalized therapies in infectious disease. Finally, the prediction of adaptive outcome can lead to predictions of future cross-resistance or collateral sensitivities. Overall, this body of work exemplifies how computational methods can complement the increasingly rapid data generation in the lab, and pave the way to the development of more effective antibiotic stewardship practices.