minghu song university of connecticut
Deep generative models have recently emerged as encouraging tools for the de novo molecular structure generation. Even though considerable advances have been achieved in recent years, the field of generative molecular design is still in its infancy. One potential solution may be to integrate domain knowledge of structural or medicinal chemistry into the data-driven machine learning process to address specific deep molecule generation goals. This manuscript proposes a new graph-based hierarchical variational autoencoder (VAE) model for molecular generation. Training molecules are first decomposed into small molecular fragments. Unlike other motif-based molecular graph generative models, we further group decomposed fragments into different interchangeable fragment clusters according to their local structural environment around the attachment points where the bond-breaking occurs. In this way, each chemical structure can be transformed into a three-layer graph, in which individual atoms, decomposed fragments, or obtained fragment clusters act as graph nodes at each corresponding layer, respectively. We construct a hierarchical VAE model to learn such three-layer hierarchical graph representations of chemical structures in a fine-to-coarse order, in which atoms, decomposed fragments, and related fragment clusters act as graph nodes at each corresponding graph layer. The decoder component is designed to iteratively select a fragment out of a predicted fragment cluster vocabulary and then attach it to the preceding substructure. The newly introduced third graph layer will allow us to incorporate specific chemical structural knowledge, e.g., interchangeable fragments sharing similar local chemical environments or bioisosteres derived from matched molecular pair analysis information, into the molecular generation process. It will increase the odds of assembling new chemical moieties absent in the original training set and enhance structural diversity/novelty scores of generated structures. Our proposed approach demonstrates comparatively good performance in terms of model efficiency and other molecular evaluation metrics when compared with several other graph- and SMILES-based generative molecular models. We also analyze how our generative models' performance varies when choosing different fragment sampling techniques and radius parameters that determine the local structural environment of interchangeable fragment clusters. Hopefully, our multi-level hierarchical VAE prototyping model might promote more sophisticated works of knowledge-augmented deep molecular generation in the future.
download asset MRGVAE_Manuscript.pdf 3 MB [opens in a new tab] cloud_download
pdf : 3 MB
download asset 28217118.pdf 8 MB [opens in a new tab] cloud_download
pdf : 8 MB