Building a Physically Interpretable Compressed Chemical Space via Grid-Based Optimization

Abstract

Abstract In recent years, deep generative models have been increasingly adopted for molecular optimization and design. In many such models, molecules are encoded into latent spaces that provide compact representations of their features. However, these latent spaces remain difficult to interpret in physical terms, and it is unclear how much structural and molecular information is retained under compression. This limitation becomes especially important when molecular behavior depends on specific three-dimensional arrangements embedded in an environmental context. Here, we introduce COMPRESS (Compression Of Molecular Physical fields into Reduced Spatial Sites), an optimization framework that represents a small molecule using a fixed number of physically parametrized spatial sites. Each site is defined by three spatial coordinates and three interaction parameters, yielding a fixed-size representation that remains directly interpretable in physical terms. To preserve molecular information, COMPRESS representations are optimized to reproduce the density, electrostatic, and van der Waals fields of the corresponding all-atom molecule on discretized three-dimensional grids. We systematically evaluate how molecular information is retained as the number of sites, K, is reduced. Our analysis covers benzene, 20 chemically diverse chemical scaffolds, benzene dimers, protein–ligand complexes, and 10,000 molecules chosen from the GEOM-Drugs data set. Across systems, field agreement deteriorates slowly at high resolution and then undergoes a rapid, pronounced transition at lower resolution. We further find that K = 25 conformers capture about 97% of the molecules from GEOM-Drugs under a loss threshold of 0.05, which we use as an indicator of fidelity. These results show that COMPRESS can provide a fixed-size, physically interpretable molecular representation, enabling compression to be evaluated directly using molecular interaction fields and resulting physical quantities. We also introduce a decoder that reconstructs all-atom structures directly from a COMPRESS representation, showing that the representation can support molecular generation in addition to field-preservation analysis. Finally, rationally modified COMPRESS representations through changes to site charge and size or the addition or removal of sites are shown to generate realistic chemical compounds upon decoding.