Generative models for structure-based molecular design hold considerable promise for drug discovery, with the potential to speed up the hit-to-lead development cycle while improving the quality of drug candidates and reducing costs. Data sparsity and bias are, however, the two main roadblocks to the development of three-dimensionally aware models. Here we propose a training protocol based on multilevel self-contrastive learning for improved bias control and data efficiency. The framework leverages the large data resources available for two-dimensional generative modelling with datasets of ligand–protein complexes, resulting in hierarchical generative models that are topologically unbiased, explainable and customizable. We show how, by deconvolving the generative posterior into chemical, topological and structural context factors, we not only avoid common pitfalls in the design and evaluation of generative models, but also gain detailed insight into the generative process itself. This improved transparency considerably aids method development and allows fine-grained control over novelty versus familiarity.