Learning metabolic dynamics from irregular observations by Bidirectional Time-Series State Transfer Network.

Area of Science:

• Biotechnology
• Systems Biology
• Machine Learning

Background:

• Modeling microbial metabolic dynamics is crucial for optimizing biomanufacturing processes.
• Traditional white-box models struggle with industrial strains lacking detailed network information.
• Existing black-box models often require pre-processing of irregular, time-series data, introducing errors.

Purpose of the Study:

• To introduce a novel deep learning approach for modeling metabolic dynamics directly from irregular time-series observations.
• To address the limitations of existing methods in handling real-world biomanufacturing data, which is often noisy and incomplete.
• To improve the accuracy and robustness of metabolic dynamic modeling for industrial applications.

Main Methods:

• Development of the Bidirectional Time-Series State Transfer Network (BTSTN), a deep learning architecture.
• Training and evaluation using synthetic datasets from ideal dynamic systems.
• Validation with a real-world fermentation process dataset exhibiting irregular measurements and noise.

Main Results:

• BTSTN accurately reconstructs dynamic behaviors of microbial metabolic systems.
• The model demonstrates robust prediction of future metabolic trajectories.
• BTSTN outperforms state-of-the-art methods in handling missing measurements and noise.

Conclusions:

• BTSTN offers a powerful new tool for modeling microbial metabolic dynamics, particularly from challenging, irregular data.
• This approach enhances the reliability and accuracy of biomanufacturing process optimization.
• BTSTN's ability to natively learn from irregular data makes it a significant advancement in the field.