Computational Framework for Machine-Learning-Enabled ¹³C Fluxomics.

Area of Science:

• Synthetic Biology
• Metabolic Engineering
• Computational Biology

Background:

• 13C metabolic flux analysis (MFA) is crucial for synthetic biology but is computationally intensive and sensitive to initial conditions.
• Optimization-based MFA often yields long computation times and unstable solutions.

Purpose of the Study:

• To develop a machine learning (ML) framework to enhance the speed and stability of 13C fluxomics.
• To create a reliable and high-throughput method for metabolic phenotyping.

Main Methods:

• Generated training and test datasets using metabolic network decomposition and flux sampling.
• Utilized flux ratios as training targets and metabolite labeling patterns as features.
• Implemented automated processes for flux ratio selection and feature screening based on importance and solvability.

Main Results:

• The ML framework significantly improved prediction accuracy by incorporating both amino acids and central carbon metabolites.
• Predicted flux ratios, combined with external fluxes, accurately determined global flux distributions.
• Validated the approach using simulated and experimental data, showing comparable or superior performance to canonical 13C MFA.

Conclusions:

• The developed ML framework offers a reliable and efficient alternative to traditional 13C MFA.
• This approach facilitates high-throughput metabolic phenotyping and advances the application of intelligent algorithms in synthetic biology's Test and Learn phase.