Molecular docking is one of the most widely used techniques for virtual screening (VS) of potential drug candidates. Despite its popularity, docking accuracy is often limited due to the trade-off between speed and precision required for screening large compound libraries. In the present work, we leverage graph convolutional networks (GCNs), a state-of-the-art deep neural network architecture, to enhance docking capacity for prioritizing active compounds from a library of ∼200,000 compounds screened against Cruzain. We propose strategies to integrate both techniques into a single VS pipeline. By applying the GCN as a pre-docking filter, the compound library was enriched with active molecules, resulting in higher hit rates in subsequent docking screenings. Additionally, to further enhance the docking performance, the GCN-learned atomic embeddings were directly incorporated into the docking process through pharmacophoric restraints. Unlike common approaches that use deep learning (DL) scoring functions to rank pre-generated docking poses, the approaches we propose here have the advantage that only compounds that passed the DL filters need to be screened by the more computationally demanding docking method. This work might serve as a proof of concept for combining deep learning and classical docking in drug discovery.