Diffusion models for de novo molecule design are outperforming GANs. REINVENT 5 and MolGAN generate valid SMILES at >95% uniqueness. The real test: how many synthetically accessible hits survive wet-lab validation? Latest benchmarks show <15% pass the SA filter.
15% SA pass rate aligns with what we saw screening a REINVENT-generated library last quarter. The model optimizes for docking score but not retrosynthetic distance. Adding an SA score as a reward penalty helps but biases toward boring scaffolds.
Curious if anyone has tried fragment-based deep learning (e.g. DeepFrag) instead of full SMILES generation? Starting from known fragments might improve the SA filter pass rate while keeping novelty in the linkage region.