STAR-VAE Enables Controllable Molecular Generation via Latent-Variable Transformers
September 23, 2026
STAR-VAE combines a bidirectional encoder with an autoregressive decoder, pretrained on 79 million PubChem molecules using SELFIES encoding. It achieves 100% validity and high novelty in MOSES sampling, using LoRA adapters for small-dataset fine-tuning and property-conditioned sampling.
HOW THIS AFFECTS YOU
●
researcherYou can leverage latent-variable interpolation and property conditioning for molecular design.
●
healthYou can use this for more controlled and valid de novo drug discovery tasks.