Leveraging Systems-level Metabolic Modeling and Machine Learning to Optimize Antibiotic Combination Therapy Design

Abstract

Over the past century, the rise of antibiotic resistance (AR) has closely followed the discovery of new antibiotics and has continued to increase as antibiotic development stalled within the past few decades. Combination therapy, which involves the prescription of two or more therapeutic agents, is a promising solution for combating AR. However, the process of designing effective combination therapies is plagued by several challenges, including: (a) the exponential explosion in the combinatorial space to search as the number of drugs and dosage levels to screen increases, (b) the heterogeneity in bacterial drug response due to differences in genetic and phenotypic states, and (c) the limited mechanistic insight that empirical methods for combination therapy design currently offer. For my dissertation research, I sought to optimize the design of antibiotic combination therapies by considering intrinsic and extrinsic factors that influence the bacterial response to drug treatment (Chapter 1). To this end, I developed two computational methods that predict drug interaction outcomes (e.g., synergy) in specific cell states and growth conditions.

The first approach is CARAMeL (Chapter 2), which stands for Condition-specific Antibiotic Regimen Assessment using Mechanistic Learning. CARAMeL leverages genome-scale metabolic models and machine learning to generate condition-specific drug interaction outcome predictions. Not only is CARAMeL better at generating accurate predictions for drug interaction outcomes against Escherichia coli and Mycobacterium tuberculosis compared to previous methods, but this approach is also the first to predict single-cell, media-specific, and sequential drug interaction outcomes. By evaluating how the outcome for individual drug combinations may vary across different conditions, CARAMeL can identify drug combinations that are predicted to retain synergy regardless of fluctuations within the cell or the surrounding environment.

The second approach that I developed is called TACTIC (Chapter 3), which stands for Transfer learning And Crowdsourcing to predict Therapeutic Interactions Cross-species. TACTIC implements crowdsourcing and transfer learning to generate strain-specific drug interaction outcome predictions, which is accomplished by extending information between multiple bacteria based on genes that are orthologous between one another. Using drug interaction data measured across 12 phylogenetically diverse bacteria, I show that TACTIC can better predict drug interaction outcomes for unseen microbes compared to INDIGO (INferring Drug Interactions using chemo-Genomics and Orthology), a prior computational approach that serves as the foundation for TACTIC. With the ability to predict strain-specific drug interaction outcomes, I apply TACTIC to determine drug combinations that are predicted to have narrow-spectrum synergy; that is, selective synergistic outcomes against pathogenic (and not commensal) bacteria.

In Chapter 4, I demonstrate how CARAMeL and TACTIC can be leveraged to guide the design of clinically relevant combination therapies for the treatment of tuberculosis, one of the deadliest infectious diseases, and endophthalmitis, a serious eye infection that leads to blindness if improperly treated. Beyond the scope of the present work, CARAMeL and TACTIC hold the potential to aid the discovery of novel combination therapies with precise efficacy against other infectious diseases. To this end, both approaches are publicly available in adaptable formats that are primed for extended use by other research groups to optimize the design of antibiotic combination therapies. In the long-term future, CARAMeL and TACTIC could be extended to guide the design of combination therapies that are urgently needed outside of bacterial infections, including but not limited to fungal infections and cancer (Chapter 5).