The Anatomical Therapeutic Chemical (ATC) classification system plays a vital role in managing drug information and guiding pharmacotherapy decisions. In this study, we propose a machine learning-based framework that uses graph convolutional networks (GCNs) to predict ATC codes in 14 main therapeutic classes. Using a dataset of 5,498 drug molecules, each represented by a SMILES (Simplified Molecular Input Line Entry System) string, we employed Extended Connectivity Fingerprints (ECFPs) to encode structural information. The resulting molecular graphs were processed by a GCN model trained to handle multi-label classification. The model achieved state-of-the-art precision, with the highest performance observed in class H (systemic hormonal preparations, 0.95), class R (respiratory system, 0.87) and classes B and V (both 0.84). Other classes, such as C, G, L, and N, also exhibited strong prediction scores. This highlights the GCN’s ability to generalize across diverse molecular patterns. In addition to conventional pharmaceuticals, we emphasize the framework’s applicability to natural products, such as seaweed-derived bioactives. These compounds, rich in polysaccharides, fatty acids, and antioxidants, are increasingly explored for therapeutic use and benefit from structured classification systems for regulatory and commercial integration. Our findings support the viability of GCN-based cheminformatics for both established drugs and emerging compounds.