Diffusion models for molecule design: a 2020–2026 scoping review finds conditioning is a design variable and evaluation is still unstandardised
- A PRISMA-ScR scoping review of diffusion-based networks for de novo drug design published between 2020 and 2026.
- Diffusion did not replace earlier architectures; graph networks, autoencoders and transformers remain the backbones, with the training objective being what changed.
- Conditioning approaches fall into structural anchoring, target-aware biochemical conditioning, property-driven steering and multimodal latent conditioning; when the signal enters the reverse trajectory is itself a design choice.
- Datasets, metrics and definitions differ across papers, leaving gaps in controllability and interpretability.
💡 Points Based on the abstract
The useful claim is the third finding: with no common benchmark or metric definitions, papers cannot be ranked, so the headline validity and novelty numbers in generative papers should be read as self-reported. When you pick a generator for a project, ignore the leaderboard and test two or three on your own target with the same docking and synthesizability filters. Reviews like this are most valuable as a map of which conditioning approach fits which input you have.