MoleculeGen

AI Drug Molecule Generator

A Variational Autoencoder over SMILES strings that learns a continuous latent space of drug-like molecules, enabling property-conditioned generation and latent-space optimization.

Methods & contributions

• SMILES VAE — bidirectional GRU encoder/decoder with the reparameterization trick over a 128-dim latent space.
• Drug-likeness filters — Lipinski, Veber, Ghose, and lead-likeness rules with QED scoring and PAINS alerts.
• Latent optimization — interpolation and property-conditioned sampling for lead generation.
• Cheminformatics — RDKit descriptors, Morgan/MACCS fingerprints, and Tanimoto similarity.

Data flow

SMILES → tokenizer → GRU encoder → z ~ N(μ,σ)
      → GRU decoder → SMILES  (ELBO = recon + β·KL)

Technologies