Data-driven interpretation of mixed-culture polyhydroxyalkanoate accumulation using dual-target machine learning...

Polyhydroxyalkanoate (PHA) accumulation in mixed-culture feast-famine systems reflects nonlinear interactions among nutrient availability, biomass retention, and reactor operation, making it difficult to distinguish the factors associated with intracellular storage. Conventional single-target models based on end-of-feast PHA content may combine inherited metabolic state with current operating effects. To address this limitation, a dual-target framework was developed using PHA_max, representing end-of-feast storage state, and PHA_gain, representing within-cycle accumulation. A dataset of 1,205 cycle-level observations from three sequencing batch reactor experiments was integrated, and six machine-learning models were evaluated using leakage-safe schemes that withheld entire reactors, stages, or experiments. The pretrained tabular foundation model TabPFN showed the most consistent performance without dataset-specific tuning, achieving R² values of 0.67 for PHA_max and 0.58 for PHA_gain under blocked temporal validation. Feature attribution indicated that PHA_max was associated mainly with inherited storage state together with substrate supply and nutrient regime, whereas PHA_gain was more closely associated with biomass state, its retention, and the applied organic load. Their relative predictability varied with the validation scheme: PHA_max performed better when reactor-specific structure was represented in training, while PHA_gain retained comparatively greater predictive value when reactors or experiments were withheld. Separating absolute storage state from within-cycle accumulation therefore provides a more differentiated basis for interpreting mixed-culture PHA processes and identifies PHA_gain as a potential target for future operational guidance.