We developed a suite of models using deep learning to make hindcast predictions of the 7-day average backward-looking nitrate concentration at 46 predominantly agricultural sites across the midwestern and eastern United States. The models used daily observations of discharge and meteorological variables and static watershed attributes describing anthropogenic modification to hydrology, nitrogen application, climate, groundwater, land use and land cover, watershed physical attributes, and soils. Across all sites, discharge and watershed soil and physiographic attributes show a particularly strong influence on model performance. An analysis of drivers across sites revealed considerable regional differences related to controlling processes such as groundwater contributions. We tested several ways to pool data across sites to develop accurate models and make the most effective use of available data. Single-site models, in which models are trained and tested at a single location, showed generally strong predictive performance (median Kling-Gupta Efficiency = 0.66), and accuracy at poorly performing sites could be improved by grouping sites with similar characteristics. Developing a single model for all sites reduced performance at several locations with distinct characteristics, suggesting that there is a threshold of dissimilarity beyond which more data does not improve the model. While many deep learning studies have shown that national or even global models can outperform local models, it is not clear that this is true for water quality constituents. This study demonstrates how existing data can be combined effectively, using deep learning to develop accurate and interpretable models of instream nitrate at sites where varying processes are responsible for changes in nitrate concentration. This release provides code and data for running a suite of machine learning model to predict in stream nitrate concentration and using explainable AI to analyze model outputs and compare among modeling approaches.