Current machine learning-based solutions to model protein-ligand interactions lack the level of interpretability that physics-based methods usually provide. Here, a workflow to embed protein binding surfaces as graphs was developed to serve as a viable data structure to be processed by geometric deep learning. The developed architecture based on mixture density models was employed to accurately estimate the position and conformation of the small molecule within the binding region. The likelihood-based scoring function was compared against existing physics-based alternatives, and significant performance improvements in terms of docking power, screening power and reverse screening power were observed. Taken together, the developed framework provides a platform for utilising geometric deep-learning models for interpretable prediction of protein-ligand interactions at a residue level.