Machine learning framework for assessment of microbial factory performance

Area of Science:

• Metabolic Engineering
• Computational Biology
• Synthetic Biology

Background:

• Metabolic models predict microbial yields but struggle with real-world performance under suboptimal conditions.
• Machine learning (ML) offers complementary predictive power but requires extensive, curated datasets, which are challenging to generate for metabolic engineering.
• Existing data for microbial bio-production is often sparse, non-standardized, and prone to human error.

Purpose of the Study:

• To develop and validate a hybrid framework integrating genome-scale metabolic models (GSMMs) with data-driven methods for enhanced prediction of microbial bio-production.
• To overcome limitations of solely relying on metabolic models or ML by combining their strengths for predicting yield, titer, and rate.
• To create a robust system for assessing microbial factory performance using curated experimental data and GSMM simulations.

Main Methods:

• Manually curated a dataset of ~1200 engineered E. coli cell factories from ~100 publications.
• Augmented experimental data with features derived from iML1515 GSMM simulations under matched experimental constraints.
• Employed ensemble learning (stacked regressors: SVM, gradient boosted trees, neural networks) and data augmentation to handle data sparsity and variability.
• Utilized multiple correspondence analysis/principal component analysis to identify key factors influencing bio-production.

Main Results:

• The hybrid framework achieved high cross-validation accuracy for predicting E. coli bio-production metrics.
• Pearson correlation coefficients between 0.8 and 0.93 were obtained on unseen data, demonstrating robust predictive capability.
• Identified influential design features and bioprocess variables impacting microbial factory performance.

Conclusions:

• The integrated metabolic modeling and machine learning approach effectively predicts microbial bio-production performance (yield, titer, rate).
• This hybrid framework addresses challenges posed by sparse, non-standardized data in metabolic engineering.
• The method provides a powerful tool for optimizing microbial cell factories and accelerating bioprocess development.