Investigating the natural product subspace within the Transformer-VAE foundation model’s drug-like molecule space

Abstract

We explore the intricate world of natural product chemistry through the lens of computational modelling. Utilising a Transformer-VAE foundation model originally pre-trained on the GuacaMol dataset, known for its comprehensive collection of small drug-like molecules. We explore the COCONUT dataset’s natural products within this model’s latent space. This approach allows us to investigate these complex natural compounds’ structural organisation and relationships in a latent space tailored to smaller, drug-like molecules. Our findings provide insightful revelations about the similarities and divergences between these two distinct molecular realms. We uncover new perspectives on molecular similarity and potential bioactivity by examining how natural products, with their diverse and often complex structures, are represented and structured in a latent space initially trained on more simplistic molecules. This research sheds light on the capabilities and adaptability of pre-trained models in chemical informatics. It could help pave the way for innovative approaches in discovering and analysing natural products for pharmaceutical applications.