Using Machine Learning to Design Effective Antimicrobial Dosing Regimens

Resistant bacterial infections remain a major clinical challenge, often necessitating combination therapy, namely use of two or more antibiotics with different mechanisms of action. However, the systematic design of such therapies is still lacking. This study evaluates machine learning (ML) models as alternatives to heuristic approaches within a previously developed framework for predicting the efficacy of antibiotic combinations under clinically relevant pharmacokinetic conditions. That framework uses longitudinal data on bacterial load, captured through optical density measurements during time-kill assays, and fits a mathematical model to the data. The outcome of this exercise is a set of values for the kill rate induced on the most resistant (least susceptible) bacteria of a bacterial population exposed to antibiotics. The contribution of this work is in feeding these kill rate data points to eight different machine learning modeling methods and comparing predictions both to prior experimental test outcomes and results from heuristic modeling. The study focused on the bacterial pathogen Acinetobacter baumannii and two antibiotic pairings: ceftazidime/amikacin and ceftazidime/avibactam, tested under both synchronous and asynchronous dosing schedules. The machine-learning models, even though they differed quantitatively from each other, produced outcomes in agreement with prior work, both experimental and computational. The value of using machine learning lies in that it provides quantitative predictions, is easy to use, affords widespread availability of related software, and needs little customization. The results suggest a useful addition to our efforts for enhancing the systematic design of combination therapies against resistant pathogens.