In this project, we present a general framework that applies 3D convolutional neural networks (3DCNN) to structure-based protein functional site detection. The framework does not rely on engineered features and can extract task-dependent features automatically from the raw atom distributions. Our deep 3DCNNs achieve an average prediction recall of 0.955 at the precision threshold of 0.99, outperforming SVM and Naïve Bayes models that employ conventional features. Importantly, the 3DCNN models showed superior performance on challenging cases where 1D sequence motifs are absent but a function is known to exist. Finally, we inspect the individual contributions of each atom to the classification decisions and show that our models successfully recapitulate known 3D features about protein functional sites.