Nutrient leaching from agricultural lands contributes significantly to water pollution and eutrophication in the Upper Mississippi River Basin (UMRB), posing substantial environmental challenges. Current water quality assessment methods, either process-based or data-driven modeling, often face significant uncertainties. For example, process-based models struggle to seamlessly integrate observations to update model structures or parameters for improved prediction, while most existing machine learning approaches primarily focus on temporal predictions, failing to capture the spatial complexity of nitrate transport due to sparse observational data and inadequate modeling approaches. This study introduces a knowledge-guided graph machine learning (KGML-Graph) approach, which combines advanced process-based modeling with graph neural network, to predict discharge and nitrate load at the HUC12 watershed scale in UMRB. A synthetic dataset covering the period from 2000 to 2020 was generated using the SWAT+ model, which incorporates meteorological, soil, topography, and land use data to simulate watershed hydrology and nutrient transport. The KGML-Graph model was initially pretrained on the synthetic data and subsequently fine-tuned with real-world observations from the United States Geological Survey. The model was tested in both temporal and spatial extrapolation scenarios to evaluate its robustness. Results demonstrate that the graph-based model significantly outperforms temporal deep learning models, such as Long Short-Term Memory (LSTM), achieving improved agreement with observed nitrate levels. The KGML-Graph model effectively captures spatial variability and generalizes across watersheds, with Kling-Gupta Efficiency (KGE) and Nash-Sutcliffe Efficiency (NSE) values of approximately 0.7 in selected watersheds. Daily nitrate load dynamics were successfully mapped across different watersheds, showing consistent performance even in data-scarce regions. These findings highlight the potential of the KGML-Graph approach for practical applications in watershed management. The proposed modeling framework offers valuable insights for improving water quality monitoring and supports the development of sustainable agricultural and conservation strategies.