Abstract

Saccharomyces cerevisiae is an important microbial organism in industrial biotechnology and metabolic engineering, known for its genetic flexibility, well-characterised metabolism, and fermentative capacity. Beyond biofuel applications, yeast is increasingly recognised as a key source for single-cell protein (SCP) production, a rising solution to address global protein-supply challenges. This study presents a comprehensive computational framework combining the Yeast9 consensus genome-scale metabolic model (GEM) with machine learning and optimisation techniques to systematically predict, interpret, and enhance biomass flux for SCP yield. The Yeast9 GEM, comprising 4,131 reactions, 2,806 metabolites, and 1,161 genes, was simulated using flux balance analysis (FBA) by varying glucose, oxygen, and ammonium uptake rates, generating 2,000 flux profiles. These profiles trained Random Forest and XGBoost regressors achieving R₂ values of 0.9999760 and 0.9997702, respectively. A variational autoencoder (VAE) with KL annealing and gradient clipping identified four metabolic clusters in latent space (silhouette peak at k=3, score = 0.334; k=4 selected for interpretability), with mean biomass fluxes of 0.472, 0.493, 0.527, and 0.505 gDW·hr−1. SHAP-based feature attribution identified twenty key reactions contributing to glycolysis, the TCA cycle, and amino-acid biosynthesis relevant to SCP nutritional quality. In silico knockout of these reactions confirmed that 18/20 (90%) are essential for growth. Bayesian optimisation of nutrient uptake conditions using scikit-optimize produced a 12.13-fold improvement in predicted biomass flux (0.0858 to 1.041 gDW·hr−1) at optimal conditions of glucose = −20.0, oxygen = −20.0, and ammonium = −8.9 mmol·gDW−1·hr−1. A generative adversarial network (GAN) generated novel flux configurations with a variance of 0.124; stoichiometric feasibility verification returned 0/100 feasible profiles (0%), attributable to incomplete generator convergence (Gloss = 1.619 vs. ideal 0.693), reported as a methodological limitation. A Pareto front analysis revealed an optimal SCP operating point at 0.0858 gDW·hr−1 biomass flux with a total amino-acid biosynthesis score of 1000.029 mmol·gDW−1·hr−1.

Abstract

Saccharomyces cerevisiae is an important microbial organism in industrial biotechnology and metabolic engineering, known for its genetic flexibility, well-characterised metabolism, and fermentative capacity. Beyond biofuel applications, yeast is increasingly recognised as a key source for single-cell protein (SCP) production, a rising solution to address global protein-supply challenges. This study presents a comprehensive computational framework combining the Yeast9 consensus genome-scale metabolic model (GEM) with machine learning and optimisation techniques to systematically predict, interpret, and enhance biomass flux for SCP yield. The Yeast9 GEM, comprising 4,131 reactions, 2,806 metabolites, and 1,161 genes, was simulated using flux balance analysis (FBA) by varying glucose, oxygen, and ammonium uptake rates, generating 2,000 flux profiles. These profiles trained Random Forest and XGBoost regressors achieving R₂ values of 0.9999760 and 0.9997702, respectively. A variational autoencoder (VAE) with KL annealing and gradient clipping identified four metabolic clusters in latent space (silhouette peak at k=3, score = 0.334; k=4 selected for interpretability), with mean biomass fluxes of 0.472, 0.493, 0.527, and 0.505 gDW·hr−1. SHAP-based feature attribution identified twenty key reactions contributing to glycolysis, the TCA cycle, and amino-acid biosynthesis relevant to SCP nutritional quality. In silico knockout of these reactions confirmed that 18/20 (90%) are essential for growth. Bayesian optimisation of nutrient uptake conditions using scikit-optimize produced a 12.13-fold improvement in predicted biomass flux (0.0858 to 1.041 gDW·hr−1) at optimal conditions of glucose = −20.0, oxygen = −20.0, and ammonium = −8.9 mmol·gDW−1·hr−1. A generative adversarial network (GAN) generated novel flux configurations with a variance of 0.124; stoichiometric feasibility verification returned 0/100 feasible profiles (0%), attributable to incomplete generator convergence (Gloss = 1.619 vs. ideal 0.693), reported as a methodological limitation. A Pareto front analysis revealed an optimal SCP operating point at 0.0858 gDW·hr−1 biomass flux with a total amino-acid biosynthesis score of 1000.029 mmol·gDW−1·hr−1.

Abstract

Saccharomyces cerevisiae is an important microbial organism in industrial biotechnology and metabolic engineering, known for its genetic flexibility, well-characterised metabolism, and fermentative capacity. Beyond biofuel applications, yeast is increasingly recognised as a key source for single-cell protein (SCP) production, a rising solution to address global protein-supply challenges. This study presents a comprehensive computational framework combining the Yeast9 consensus genome-scale metabolic model (GEM) with machine learning and optimisation techniques to systematically predict, interpret, and enhance biomass flux for SCP yield. The Yeast9 GEM, comprising 4,131 reactions, 2,806 metabolites, and 1,161 genes, was simulated using flux balance analysis (FBA) by varying glucose, oxygen, and ammonium uptake rates, generating 2,000 flux profiles. These profiles trained Random Forest and XGBoost regressors achieving R₂ values of 0.9999760 and 0.9997702, respectively. A variational autoencoder (VAE) with KL annealing and gradient clipping identified four metabolic clusters in latent space (silhouette peak at k=3, score = 0.334; k=4 selected for interpretability), with mean biomass fluxes of 0.472, 0.493, 0.527, and 0.505 gDW·hr−1. SHAP-based feature attribution identified twenty key reactions contributing to glycolysis, the TCA cycle, and amino-acid biosynthesis relevant to SCP nutritional quality. In silico knockout of these reactions confirmed that 18/20 (90%) are essential for growth. Bayesian optimisation of nutrient uptake conditions using scikit-optimize produced a 12.13-fold improvement in predicted biomass flux (0.0858 to 1.041 gDW·hr−1) at optimal conditions of glucose = −20.0, oxygen = −20.0, and ammonium = −8.9 mmol·gDW−1·hr−1. A generative adversarial network (GAN) generated novel flux configurations with a variance of 0.124; stoichiometric feasibility verification returned 0/100 feasible profiles (0%), attributable to incomplete generator convergence (Gloss = 1.619 vs. ideal 0.693), reported as a methodological limitation. A Pareto front analysis revealed an optimal SCP operating point at 0.0858 gDW·hr−1 biomass flux with a total amino-acid biosynthesis score of 1000.029 mmol·gDW−1·hr−1.

Abstract

Saccharomyces cerevisiae is an important microbial organism in industrial biotechnology and metabolic engineering, known for its genetic flexibility, well-characterised metabolism, and fermentative capacity. Beyond biofuel applications, yeast is increasingly recognised as a key source for single-cell protein (SCP) production, a rising solution to address global protein-supply challenges. This study presents a comprehensive computational framework combining the Yeast9 consensus genome-scale metabolic model (GEM) with machine learning and optimisation techniques to systematically predict, interpret, and enhance biomass flux for SCP yield. The Yeast9 GEM, comprising 4,131 reactions, 2,806 metabolites, and 1,161 genes, was simulated using flux balance analysis (FBA) by varying glucose, oxygen, and ammonium uptake rates, generating 2,000 flux profiles. These profiles trained Random Forest and XGBoost regressors achieving R₂ values of 0.9999760 and 0.9997702, respectively. A variational autoencoder (VAE) with KL annealing and gradient clipping identified four metabolic clusters in latent space (silhouette peak at k=3, score = 0.334; k=4 selected for interpretability), with mean biomass fluxes of 0.472, 0.493, 0.527, and 0.505 gDW·hr−1. SHAP-based feature attribution identified twenty key reactions contributing to glycolysis, the TCA cycle, and amino-acid biosynthesis relevant to SCP nutritional quality. In silico knockout of these reactions confirmed that 18/20 (90%) are essential for growth. Bayesian optimisation of nutrient uptake conditions using scikit-optimize produced a 12.13-fold improvement in predicted biomass flux (0.0858 to 1.041 gDW·hr−1) at optimal conditions of glucose = −20.0, oxygen = −20.0, and ammonium = −8.9 mmol·gDW−1·hr−1. A generative adversarial network (GAN) generated novel flux configurations with a variance of 0.124; stoichiometric feasibility verification returned 0/100 feasible profiles (0%), attributable to incomplete generator convergence (Gloss = 1.619 vs. ideal 0.693), reported as a methodological limitation. A Pareto front analysis revealed an optimal SCP operating point at 0.0858 gDW·hr−1 biomass flux with a total amino-acid biosynthesis score of 1000.029 mmol·gDW−1·hr−1.