Characterizing and Exploiting the Latent Space of a Graph Convolutional Network for Bioisostere Identification
Year: 2026
Type: preprint
Abstract: We show that the latent vectors, broken down to atom and fragment contributions, produced by a graph convolutional network (GCN) encode chemically meaningful information. Latent vectors for atoms, further characterized by their neighbors, and fragments are largely independent of the rest of the molecule. This observation extends to transformations, making the organization of the latent space comparable to the capturing of syntax and semantics in word embeddings such as word2vec. Applying this observation to bioisostere identification, we find that bioisosteric fragment pairs are also closer in latent space than random pairs. Building on this result, we developed a method to generate bioisosteric suggestions.
Source: ChemRxiv
Authors: Peter Gedeck, Sophie Bekerman, Philip P. Cheung, Alex M. Clark, Michael Bowles, Barry A. Bunin
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7141635430
Cites:
Cited by:
Related to: