Hybrid genome‑scale metabolic modeling and machine learning reveal cost‑efficient strategies for phototrophic polyhydroxybutyrate production in Rhodopseudomonas palustris.

Plastic pollution from fossil-based materials is a major global environmental challenge. Microbially derived bioplastics, such as polyhydroxybutyrate (PHB), offer a promising biodegradable alternative. However, the high substrate and operational costs of PHB production remain a major barrier to large-scale deployment. Optimizing PHB synthesis requires navigating a multidimensional design space of metabolic, nutritional, and operational variables, which is impractical to explore experimentally. This study presents an integrated computational framework that combines a genome-scale metabolic model (GEM) of Rhodopseudomonas palustris, machine-learning surrogate modeling, Pareto multi-objective optimization, and thermodynamics-based flux analysis (TFA) to identify cost-efficient and biologically feasible PHB production strategies. Experimental and literature-derived medium compositions were translated into mechanistic constraints, enabling the GEM to generate metabolically coherent synthetic datasets that augmented sparse experimental observations. CatBoost surrogate machine-learning (ML) models trained on this hybrid dataset accurately predicted PHB synthesis across thousands of hypothetical conditions, and Pareto optimization revealed operating regimes that balance PHB productivity with nutrient cost. TFA validated the thermodynamic feasibility of these strategies and refined pathway usage, reinforcing thiolase-initiated routing into PHB biosynthesis and suppressing infeasible β-oxidation-like redox loops. Overall, this hybrid GEM-ML-TFA framework identifies metabolic bottlenecks, engineering targets, and cost-optimal nutrient regimes for phototrophic PHB production, providing a scalable blueprint for rational process and strain design.