Warped Latent Spaces and Traversal for Chemical Deep Generative Models

ICLR 2026 Conference Submission22019 Authors

Keywords: deep generative models, chemical deep generative models, latent variable models, variational autoencoders, property prediction, drug design

TL;DR: Learning and traveral in chemically informed latent spaces of deep generative models for small molecule generation

Abstract: We propose a generative framework for interpretable and property-aware molecular design by learning warped subspaces within the latent space of a chemical variational autoencoder (VAE) trained on a sequential representation of small molecules. Instead of directly regularizing latent coordinates, our approach works by creating low dimensional subspaces that are smoothly warped to align with molecular property variation using a novel alignment loss. This warping provides a flexible mechanism to capture nonlinear structure in property–latent relationships while retaining interpretability. This framework enables property optimisation and traversal within a low-dimensional subspace, where directions correspond to meaningful variations in molecular properties and decode back into valid molecules in the original space. We evaluate the method on various tasks related to conditional molecular generation on standard benchmarks used in literature like QM9, ZINC250K and the Pubchem drug datasets demonstrating strong generative quality, validity, uniqueness and novelty alongside a more controllable approach molecular generation.

Primary Area: generative models

Submission Number: 22019