Non-linearity of Metabolic Pathways Critically Influences the Choice of Machine Learning Model.

Area of Science:

• Computational Biology
• Systems Biology
• Machine Learning Applications

Background:

• Machine learning (ML) accelerates model development in life sciences.
• Mechanistic models and metabolic networks are established tools for biological pathway design.
• The application of ML to metabolic pathway modeling and their inherent non-linearity remains underexplored.

Purpose of the Study:

• To construct and evaluate various linear and non-linear ML models for metabolic pathway analysis.
• To assess the impact of data features on model performance in predicting biological data.
• To establish decision-making support for selecting appropriate models in metabolic pathway research.

Main Methods:

• Development of metabolic pathways using diverse datasets.
• Implementation of multiple linear and non-linear ML algorithms (e.g., Quantile Random Forest, Bayesian Generalized Linear Model).
• Comparative analysis of model performance based on prediction accuracy for pathway flux and product concentration.

Main Results:

• Non-linear ML models demonstrate superior performance compared to linear models.
• Quantile Random Forest (QRF) achieved high accuracy (RMSE = 0.021 nmol·min⁻¹, R² = 1) versus Bayesian GLM (RMSE = 1.379 nmol·min⁻¹, R² = 0.823).
• Model performance is significantly influenced by data features and pathway non-linearity.

Conclusions:

• Non-linear models are better suited for metabolic pathways, which exhibit significant non-linear characteristics.
• The findings support the predominance of non-linear dynamics in metabolic pathways.
• This research provides a framework for selecting appropriate ML models for pathway analysis, with implications for biomarker discovery and industrial optimization.