Antibiotic resistance is growing as a major public health concern. Yet, infections are often treated “empirically”, in absence of susceptibility measurements, leading to mismatches between prescription and infection susceptibility. We and others have done extensive work, including the use of Machine Learning (ML) tools, for algorithmically matching the prescribed antibiotic to the resistance profile of the infecting pathogen. However, antibiotics are a double-edged sword: while they help clear the current infection, they also select for future resistant pathogens that are harder to treat. Tailoring antibiotics for a given infection at the single patient level should therefore focus not just on maximizing clearance of the focal infection (“greedy” algorithm), but also on reducing the risk of future resistant infections (“look-ahead” algorithm). Urinary Tract Infections (UTIs) often persist and recur, necessitating a look-ahead treatment strategy. The majority of women will be affected by a UTI over their lifetime and, importantly, for many women infections persist and recur over months and years. This recurrent nature of UTIs and the common practice of empirical treatments necessitates thinking about treatment not as individual steps but as a long-term patient-specific strategy.
UTIs are often self-seeded from the gut microbiome, highlighting the microbiome strain-level composition as a predictive and even causal factor for infections. We thereby hypothesize that profiling the microbiome can help predict the resistance profile of infection and risk of recurrences, thereby guiding treatment that optimizes current and future efficacy. Furthermore, we hypothesize that it may be possible to manipulate the microbiome to be less prone to seeding infections, or to minimize the risk of resistance in future infections, should they occur. By understanding and exploiting the three-way interaction between a patient's microbiome, antibiotic treatment, and UTI risk, we aim to develop methods to predict and manipulate recurrent infections, leading to a personalized patient-specific strategy. We will: (1) use microbiome sequencing and phenotyping to predict the strain and resistance profile of infection; (2) use microbiome-informed reinforcement learning to optimize multi-step antibiotic treatments; (3) use in vitro selection of the microbiome to predict the potential of antibiotic treatments to manipulate the microbiome composition to reduce virulence and recurrence.
