Causal AI digital twin for bioprocess bottleneck diagnosis via metabolic flexibility and rigidification maps

Diagnosing underperforming non-model microbial cultures remains difficult. To address this, we developed an intervention-aware genome-informed digital twin with interpretable flexibility features, explainable learning, and causal-structure discovery to convert sparse anchors into actionable diagnostic artifacts. We grew Stenotrophomonas maltophilia SO-1 aerobically on acetate minimal medium as a controlled testbed and anchored the twin with standardized harvest-time growth phenotypes (OD 600 at 32 h). We then generated an intervention-labeled design space via Latin hypercube sampling (LHS), labeled regimes using the active-constraint set, and encoded intracellular states utilizing targeted flux-variability analysis (FVA) widths across 30 reactions/modules. Explainable learning via XGBoost + SHapley Additive exPlanations (SHAP) identified regime-specific signatures and tipping-like patterns consistent with flexibility collapse, while causal-structure discovery yielded a rigidification map of bottleneck hypotheses into an upstream-to-downstream cascade. Validation under acetate-stress, nutrient-limited, and oxygen-transfer conditions yielded regime-level diagnostic agreement and highlighted systematic mismatches as signals for non-stoichiometric constraints.