A Scoping Review of Diffusion-Based De Novo Molecular Generation: Architectures, Evaluation, and Open Challenges for Computational Drug Discovery

Diffusion models form the bulk of generative frameworks used in de novo molecular design, yet the literature remains fragmented across model formulations, molecular representations, and generative tasks, and inadequate systematic accounts exist of how these models are steered towards molecules with desired properties, an aspect of drug discovery that is paramount to acceptance. We present a scoping review of diffusion-based neural networks applied in de novo drug discovery between 2020 and 2026. Following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) protocol, we categorise the literature into non-mutually exclusive themes covering methodological contributions, conditioning mechanisms, molecular representations, symmetries, and multi-modal integrations, and outline the datasets and evaluation metrics used throughout. Three findings emerge. First, diffusion did not displace earlier architectures so much as absorb them: Graph Neural Networks (GNNs), Autoencoders, and Transformers persist as backbones, and what changed was the training objective. Second, conditioning strategies reduce to structural anchoring, target-aware biochemical conditioning, property-driven steering, and multimodal latent conditioning; within these, the point in the reverse trajectory at which a conditioning signal is applied is a design variable distinct from the signal itself. Third, evaluation remains unstandardised across frameworks, datasets, and definitions. We identify the resulting gaps in controllability and interpretability as the central open challenge for conditioned molecular generation and outline the direction to ease them as this would enable generation of molecules that match the preferred property profiles.