Reinforcement learning for adaptive control of phenotypically heterogeneous bacterial populations

Area of Science:

• Microbiology and Computational Biology
• Antimicrobial Resistance Research

Background:

• Bacterial populations exhibit resilience to antibiotics due to physiological heterogeneity.
• Environmental fluctuations exacerbate this heterogeneity, reducing the efficacy of fixed-dose treatments.
• Antimicrobial resistance poses a significant global health threat.

Purpose of the Study:

• To develop an adaptive treatment strategy framework using reinforcement learning (RL).
• To infer bacterial population physiological states from population-level measurements.
• To design data-driven strategies for combating antimicrobial resistance.

Main Methods:

• Implemented a reinforcement learning (RL) framework for adaptive treatment discovery.
• Utilized experimentally accessible, population-level measurements as input.
• Integrated control over nutrient availability into the treatment protocol.
• Trained the RL agent to infer hidden physiological states of bacterial populations.

Main Results:

• The RL agent successfully inferred hidden bacterial physiological states.
• The framework generated adaptive strategies effective even in novel conditions.
• Controlling nutrient availability alongside drug dynamics led to consistent population extinction.
• The RL approach outperformed adaptive protocols based solely on drug dynamics.

Conclusions:

• A novel RL framework can design effective adaptive antibiotic treatment strategies.
• Inferring physiological states and controlling nutrient availability are key to bacterial population control.
• This data-driven approach offers a powerful tool against antimicrobial resistance.