Drug metabolism studies are a critical component of the drug design process. Metabolism of some drugs can lead to diminished therapeutic efficacy or even toxicity. The stability of a drug is expressed by the atoms, called Sites of Metabolism (SOMs), which undergo structural changes when that drug interacts with a metabolizing enzyme. Computationally predicting these metabolically labile atoms early on in the drug development process will enable pharmaceutical chemists to design molecules with favorable metabolic properties. A number of in silico methods have been developed for identifying SOMs, with a recent focus on machine learning due to its computational efficiency over structural modeling. Machine learning techniques classify atoms as SOMs based on feature vector representations. Existing approaches rely upon expert knowledge and often expensive experiments to engineer fixed atom descriptors with extensive sets of experimentally-derived attributes. However, models based upon learned instead of fixed representations have proven promising in other chemoinformatics tasks. Seeing molecules as attributed graphs, where atoms correspond to nodes and bonds correspond to edges, the SOM prediction problem can be formulated as a node classification task. We compare two methods of extracting node features from molecular graphs: a standard fingerprint generation strategy used by existing SOM prediction methods, which constructs task-agnostic node descriptors, and an unexplored approach based on a graph convolutional neural network, which learns taskspecific node encodings. Both methods take into account the node attributes and graph connectivity to generate descriptive atom representations. We experiment with parameters that can influence the performance of both feature extraction methods on a dataset commonly used in the literature for predicting SOMs. Despite the fact that the graph convolution approach requires more data and has more parameters to tune, we have achieved comparable performance between the two methods. Given enough data, we believe the graph convolution approach may reliably achieve improved performance over the fingerprint generation strategy. Our results indicate that the graph convolution approach can outperform the fixed fingerprint generation strategy when starting from molecular graphs that are not initialized with rich electro-chemical properties, demonstrating how learned representations could replace the need for expert-derived features for SOM prediction. Our results also illustrate the importance of tuning the feature extraction method to the metabolizing enzyme of interest.