A machine learning framework is described for performing generation of candidate molecules for, e.g., drug discovery or other applications. The framework utilizes a pre-trained encoder-decoder model to interface between representations of molecules and embeddings for those molecules in a latent space. A fusion module is located between the encoder and decoder and is used to fuse an embedding for an input molecule with embeddings for one or more exemplary molecules selected from a database that is constructed according to a design criteria. The fused embedding is decoded using the decoder to generate a candidate molecule. The fusion module is trained to reconstruct a nearest neighbor to the input molecule from the database based on the sample of exemplary molecules. An iterative approach may be used during inference to dynamically update the database to include newly generated candidate molecules.