A mechanism-aware and multiomic machine-learning pipeline characterizes yeast cell growth

Area of Science:

• Systems and synthetic biology
• Computational biology
• Machine learning applications in biology

Background:

• Metabolic modeling and machine learning are crucial for understanding genotype-phenotype-environment relationships.
• Integrating these approaches maximizes their potential, but their combined use for omic data augmentation is underexplored.
• Predicting yeast cell growth requires integrating diverse biological data.

Purpose of the Study:

• To assess and compare machine-learning-based data integration techniques.
• To combine gene expression profiles with metabolic flux data for improved prediction of yeast cell growth.
• To develop and validate a novel multiview neural network for biological data integration.

Main Methods:

• Created 1,143 strain-specific metabolic models for *Saccharomyces cerevisiae* mutants.
• Tested 27 machine-learning methods, including feature selection and multiview learning.
• Developed a multiview neural network integrating fluxomic and transcriptomic data.

Main Results:

• The multiview neural network using fluxomic and transcriptomic data increased predictive accuracy compared to transcriptomic data alone.
• Fluxomic data revealed functional patterns not evident from gene expression alone.
• The model demonstrated robustness on an independent dataset and improved predictions even for unmodeled knockout strains.

Conclusions:

• Fusing experimental data with in silico metabolic models provides complementary information for biologically informed machine learning.
• This integration enhances prediction accuracy and the interpretability of mechanistic biological insights.
• The study offers tools for understanding and manipulating complex phenotypes in systems and synthetic biology.