Per-MOF Mechanistic Attribution Reveals Context-Dependent Stability Mechanisms in Metal-Organic Frameworks.

Water stability remains a critical barrier to the practical deployment of metal-organic frameworks (MOFs) in aqueous and industrial environments. While recent machine-learning studies have achieved strong predictive performance for MOF water stability, most rely on global feature importance rankings that implicitly assume uniform mechanisms across all frameworks. However, degradation mechanisms are known to vary with metal chemistry, linker composition, and framework topology, suggesting that instance-level analysis is needed to capture this mechanistic diversity. Here, we apply per-MOF SHAP (SHapley Additive exPlanations) analysis to the recently curated WS24 data set, comprising 1,092 experimentally characterized frameworks with water stability labels. Unlike global feature selection methods such as recursive feature addition, per-MOF SHAP decomposes each prediction into stabilizing and destabilizing contributions from individual descriptors, enabling direct attribution of model-level feature contributions for specific MOFs. Using gradient-boosted decision trees with SHAP-selected features, we achieve test set performance of ROC-AUC = 0.811 and balanced accuracy = 0.723, comparable to recursive feature addition approaches on WS24. Instance-level analysis reveals that MOF water stability arises from multiple, context-dependent descriptor patterns: metal-centered descriptor patterns associated with unfavorable electronic properties, linker-dominated hydrolysis in frameworks with vulnerable organic chemistry, and context-dependent porosity effects. Quantitative analysis demonstrates that gravimetric surface area shows positive SHAP contributions in 73% of correctly predicted stable frameworks but negative contributions in 48% of correctly predicted unstable frameworks (t = 2.15, p = 0.033), confirming that identical descriptors can produce opposite model contributions depending on the chemical environment. This work establishes a practical framework for using interpretable machine learning to guide rational MOF design for aqueous applications while transparently revealing where current models and descriptors require refinement.