Deep generative models for 3D structure-based drug design and molecular optimisation: a comprehensive survey

Structure-based drug molecule generation remains a central challenge in computer-aided drug design. Recent advances in three-dimensional (3D) deep generative models have improved molecular design capabilities; however, progress in this field is hindered by the lack of a unified cross-paradigm taxonomy and fragmented evaluation practices, particularly in the transition from diffusion-based methods to flow matching and synthesizability-aware generation. To address these limitations, this survey provides a comprehensive review of over 100 methods in 3D structure-based drug design (SBDD) and molecular optimisation. First, we propose a unified taxonomy covering four primary generative paradigms: autoregressive models, diffusion models, flow matching, and Bayesian Flow Networks. Second, we systematically analyse evaluation inconsistencies in current benchmarking practices, including geometric relaxation artefacts in docking-based metrics, and introduce a standardised geometric validity auditing perspective to improve comparability. Third, we identify four cross-cutting evolutionary trends in SBDD: SE(3)-equivariant modelling, transition from sequential to parallel generation, shift from implicit to explicit conditioning, and integration of synthesizability constraints into generative priors. Finally, we summarise seven open challenges, provide a scenario-based decision framework for practitioner method selection, and highlight SE(3)-equivariant flow matching as a promising direction for future foundation models, balancing efficiency, geometric fidelity, and multi-objective optimisation.