Marine and estuarine sediments play a critical role in the global carbon cycle, particularly in the long-term storage of dissolved organic matter (DOM). Iron (Fe), through the formation of the "iron curtain" at specific redox boundaries, facilitates the capture and preservation of DOM, especially terrestrial DOM. The interaction between Fe and DOM is primarily influenced by the molecular structure of DOM and the redox environment in which it resides, leading to variations in deposition potential and long-term preservation in sediments. Due to the significant differences in the behavior of DOM molecules during Fe precipitation, there is an urgent need for new methods to predict which DOM molecules are more likely to bind with Fe and deposit. This study develops a customized graph neural network (GNN) model to predict which DOM molecules are most likely to bind with Fe and deposit in marine and estuarine sediments. The model integrates the molecular structure of DOM with environmentally constrained features and employs deep neural networks (DNN) to predict the depositional potential of DOM molecules. Particularly for terrestrial DOM molecules, graph learning helps identify which molecules are more likely to deposit in redox boundaries, offering new insights into their long-term preservation in the global carbon cycle. This approach highlights the advantages of graph learning in processing the complex structural features of DOM and elucidating the relationship between Fe precipitation and DOM preservation, offering new insights into the role of Fe in carbon sequestration and the preservation of organic matter in sediments.