Multi-stage variational autoencoders for hierarchical molecular generation and activity optimization

Deep generative models may detect novel compounds with favourable features, exhibiting chemical design potential. Traditional single-stage variational autoencoders (VAEs) lack validity, uniqueness, and biologically meaningful distribution alignment. It is difficult to represent global molecular architecture and chemical properties in a single latent representation. To overcome these challenges, we offer a multi-stage VAE system that encodes and decodes molecular representations in sequence. Improvements to latent space retain structural integrity while also adding innovation and distinction. Validity, originality, novelty, Fréchet ChemNet Distance (FCD), and KL divergence are used to validate the methodology with ChEMBL and polymer datasets. The bioefficacy of EGFR inhibitors is evaluated using computational Chemprop-based QSAR models. We offer adaptive fine-tuning strategies for the inner-layer (IL) and outer-layer (OL) to improve generating accuracy. IL adaptability is most suited to active compounds. Quantitative evaluations indicate consistent gains in validity, novelty, and biological activity over strong baselines (for example, MoLeR and RationaleRL). We give MNIST tests that confirm the hierarchical training method's stability but not its scalability beyond molecular tasks, ensuring cross-domain applicability. For generative drug discovery, hierarchical latent models with a multi-stage VAE are advised.