STAR-VAE: Latent Variable Transformers for Scalable and Controllable Molecular Generation 1. The vast chemical space of drug-like molecules necessitates powerful generative models. STAR-VAE addresses this by combining a Transformer encoder and autoregressive Transformer decoder, trained on 79 million molecules using SELFIES to ensure syntactic validity. This approach enables both broad distribution learning and conditional generation guided by molecular properties. 2. A key innovation is the principled conditional latent-variable formulation. A property predictor provides a consistent conditioning signal to the latent prior, inference network, and decoder. This allows STAR-VAE to generate molecules with desired properties using limited labeled data, making it highly efficient for property-guided molecular design. 3. Efficiency is further enhanced through low-rank adapters (LoRA) in both the encoder and decoder. This enables fast fine-tuning with minimal data, making the model adaptable to new tasks without extensive retraining. This is crucial for practical applications where labeled data is scarce. 4. STAR-VAE demonstrates strong performance on benchmarks like GuacaMol and MOSES, matching or exceeding existing baselines. In conditional tasks, it shifts docking score distributions toward stronger binding affinities, as shown in the Tartarus protein–ligand design benchmark. This highlights its potential for drug discovery. 5. Latent space analyses reveal smooth, semantically structured embeddings, supporting both unconditional exploration and property-aware steering. This dual capability makes STAR-VAE a versatile tool for navigating the complex landscape of chemical space. 📜Paper: arxiv.org/abs/2511.02769v1 #MolecularGeneration #Transformer #VAE #DrugDiscovery #AIinChemistry