Environmentally constrained machine learning uncovers climate-driven optical-molecular coupling and turnover of dissolved organic matter.

Whether fluorescence retains molecular meaning across changing redox and hydrodynamic regimes remains unresolved, limiting interpretation of dissolved organic matter (DOM) turnover in stratified inland waters. Here, we developed an environmentally constrained framework integrating EEM-PARAFAC, FT-ICR MS, and ecological niche partitioning to resolve optical-molecular coupling. The framework distinguishes fluorescent stable molecules (FSMs) from fluorescent variable molecules (FVMs) and estimates whether molecules are associated with specific fluorescent components under defined environmental conditions. Optical-molecular coupling was environmentally contingent rather than fixed, and identical fluorescence signals corresponded to different molecular assemblages across hydrodynamic states. All identified FSMs were confined to the recalcitrant dissolved organic matter (RDOM) domain, consistent with cross-niche-stable fluorescence associations being concentrated in low-H/C molecular scaffolds. Humic-like fluorescence was linked to compositionally conserved molecular cores represented by FSMs and appeared relatively buffered across habitats, whereas protein-like fluorescence was linked exclusively to FVMs and responded more strongly to oxygen, thermal, and nutrient forcing. These results show that fluorescence components partition DOM along a stability-responsiveness spectrum, providing a data-driven basis for interpreting DOM turnover, molecular reorganization, and carbon retention while offering a scalable framework for inferring fluorescence-associated molecular assemblages where routine FT-ICR MS characterization is impractical.