Molecular Generative Models Internally Organize Chemical Identity

Generative models designed for matter, particularly molecules, are typically assessed by their ability to sample diverse outputs and by how their latent spaces facilitate navigation of chemical properties. However, less is understood about the internal mechanisms by which these models structure distinct chemical identities within their representations. This study delves into this internal organization by tracing molecular identity back through the generative process. The findings reveal that models establish a fixed, piecewise-constant partition, which dictates the range of molecular objects they can generate. The research highlights that this internal repertoire's structure is influenced by several factors, including the specific representation used, the definition of chemical identity, the degree of randomness in the decoder, and the metrics employed for comparing coordinates. During training, while local chemical organization stabilizes, the diversity of distinct molecular identities within neighborhoods continues to evolve. This suggests that the internal organization of these generative spaces must be explicitly characterized rather than assumed, especially when aiming to use them for chemically meaningful navigation.