De novo generation of novel, drug-like molecules. Set target property ranges (MW, logP, HBD/HBA) + an optional seed, and Physon.Design proposes 20 new candidates distant from the seed in fingerprint space. Every candidate is RDKit-validated and Lipinski-checked before returning.
SMILES-diffusion model (D3PM) trained on ChEMBL drug-like subset. Reverse-diffuses from an all-mask sequence to a valid molecule, classifier-free-guided by your target property vector.
Morgan-fingerprint Tanimoto distance to your seed. Score = 1 - max similarity. 0 = same as seed; 1 = completely novel scaffold.
Every returned SMILES is RDKit-validated (parseable + sanitizable), passes your property filter, and — when Ro5 mode is on — satisfies Lipinski Rule of 5.