Strain design optimization using reinforcement learning.

Area of Science:

• Synthetic biology
• Metabolic engineering
• Machine learning

Background:

• Engineered microbes offer sustainable chemical synthesis but strain optimization is inefficient, relying on costly trial-and-error.
• Current methods lack efficiency and certainty in duration and cost for optimizing microbial strains.
• Advanced techniques are needed to guide strain optimization, addressing cellular regulation complexity and limited mechanistic knowledge.

Purpose of the Study:

• To introduce a multi-agent reinforcement learning (MARL) approach for optimizing metabolic enzyme levels in microbial strains.
• To develop a model-free method that learns from experimental data to enhance production without prior knowledge of metabolic networks.
• To demonstrate the MARL approach's effectiveness in accelerating and improving the reliability of industrial-scale microbial strain engineering.

Main Methods:

• A multi-agent reinforcement learning (MARL) framework was employed to tune metabolic enzyme levels.
• The model-free MARL approach learns directly from experimental data, requiring no prior knowledge of cellular regulation.
• The method was validated using a genome-scale kinetic model of Escherichia coli (k-ecoli457) and experimental data for L-tryptophan production in Saccharomyces cerevisiae.

Main Results:

• The MARL approach demonstrated effective learning from experimental data to improve production yields.
• Performance evaluations showed favorable speed of convergence, noise tolerance, and statistical stability for practical strain engineering.
• Successful application in optimizing L-tryptophan production in yeast using public experimental data was achieved.

Conclusions:

• Multi-agent reinforcement learning (MARL) is a promising strategy for guiding microbial strain optimization.
• This approach surpasses the limitations of traditional methods by operating beyond mechanistic knowledge.
• MARL facilitates faster and more reliable development of industrially relevant microbial production strains.