Machine-Learning-Assisted Pathway Optimization in Large Combinatorial Design Spaces: A p-Coumaric Acid Case Study

Area of Science:

• Metabolic Engineering
• Synthetic Biology
• Machine Learning Applications

Background:

• Combinatorial pathway optimization enhances microbial strain performance.
• Machine learning (ML) integration in the Design-Build-Test-Learn (DBTL) cycle is promising but often limited to small design spaces.
• Exploring large combinatorial libraries is crucial for maximizing strain optimization potential.

Purpose of the Study:

• To apply ML-guided DBTL cycles for optimizing p-coumaric acid production in Saccharomyces cerevisiae.
• To evaluate a gradient bandit-based ML strategy balancing exploration and exploitation in large metabolic engineering design spaces.
• To compare the effectiveness of this balanced ML strategy against greedy and feature importance-based approaches.

Main Methods:

• Construction of a large combinatorial library with 18 genes and 20 promoters (170 million designs) for p-coumaric acid production.
• Implementation of two DBTL cycles using a gradient bandit-based ML recommendation strategy.
• Comparative analysis of the balanced ML strategy against greedy and feature importance-based methods for strain optimization.

Main Results:

• The balanced ML exploration-exploitation strategy outperformed greedy and feature importance-based approaches.
• The strategy led to increased diversity in strain performance and more effective identification of top producers.
• Application to an alternative parent strain resulted in the highest p-coumaric acid titer (1.23 g/L), a 2.37-fold improvement.

Conclusions:

• ML-guided exploration is valuable for optimizing metabolic pathways within large design spaces.
• Balancing exploration and exploitation is critical for successful strain engineering and maximizing compound production.
• The developed ML strategy offers a robust method for advancing metabolic engineering efforts.