Computationally predicting bacterial strain susceptibility to phages

About This Grant

Summary. The prevalence of antibiotic resistant (AR) bacterial infections continues to grow. Although the development of novel antimicrobial compounds is one approach to combat AR infections, phage therapy or using lytic phage to treat bacterial infections, offers another solution that has many attractive benefits, including the specificity of infection, leaving the healthy microbiome intact, low toxicity, and the diversity of phages available. However, before phage therapy can be widely used in the clinic, one of the significant challenges that must be addressed is the selection of which phage to use for a given bacterial pathogen. Phage infection specificity is complicated by the fact that bacteria encode a diverse array of phage defense systems that block phage infection, typically expressed by horizontally transferrable DNA elements. The bacterial pathogen must also encode and express the phage receptor and any bacterial host factors that the phage requires for successful replication and phage production. Currently, bacterial pathogens are manually screened against large phage biobanks to select phage cocktails that can provide effective in vivo killing. Although this has been effective, such an approach is costly and, more importantly, time-intensive, and it will be challenging to scale up as phage therapy becomes more widely used. The field, therefore, needs rapid and cost-effective approaches to identify effective phages for any bacterial pathogen, given the genome sequence of the bacteria and phages. The MPIs of this proposal, Ravi and Waters, will use their diverse expertise in bacterial pathogenesis, phage biology and defense, AR, microbial genomics, and computational biology, to develop an ML-based prediction model that can identify effective phage and phage resistance-associated molecular features for any given E. coli strain. Another critical outcome of this work will be the gold-standard data set generated in Aim 1 that will define the successful infection of 69 dsDNA E. coli phage in the well-characterized BASEL phage collection with ~600 sequenced pathogenic and non-pathogen E. coli strains generating ~42,000 unique data points. Aim 2 will first define all known phage defense, AR, and virulence elements in this collection of E. coli and merge these annotated features with the phage host infection phenotypes generated in Aim 1 using (un)supervised ML-based approaches (e.g., logistic regression, random forest) to generate models that can predict effective phage infections of any given E. coli host, along with the underlying molecular features (genes, proteins, domains) culminating in resistance/susceptibility. This model will be validated with 50 new E. coli strains. Successful completion of this proposal will generate a clinically useful predictive model for E. coli and lay the framework for generating such predictive models for phage therapies against other bacterial pathogens. Moreover, the model will lead to the discovery of novel phage defense elements and bacterial factors that impact phage infection, and a deeper understanding of how bacterial pathogens evolve resistance to phage infection, knowledge, which can be used to effectively tailor phage therapy to prevent widespread emergence of resistance.