KRONOS Generates 3D Molecules with Latent Autoregressive Diffusion

Generating three-dimensional (3D) molecules is a critical task in drug discovery and materials science, with diffusion models currently dominating due to their high generation quality. However, these models typically require the molecular size to be predefined. Autoregressive approaches offer flexibility for variable-length generation and conditioning on partial molecular contexts but have historically lagged in performance and struggled to balance unconditional and context-conditioned generation. KRONOS, a novel latent autoregressive diffusion framework, aims to bridge this gap. It operates by generating molecules within the latent space of a pre-trained autoencoder, simultaneously modeling both the molecular graph topology and its 3D geometry. This approach retains the inherent flexibility of autoregressive generation. A key innovation in KRONOS is its mixed training strategy, inspired by the Fill-in-the-Middle (FIM) paradigm. This strategy enables the single left-to-right autoregressive model to perform both unconditional and fragment-conditioned molecular generation effectively. Experiments on standard datasets like QM9 and GEOM-Drugs demonstrate that KRONOS achieves leading unconditional generation performance among autoregressive methods and remains competitive with diffusion models, all while supporting fragment-conditioned generation with minimal impact on unconditional performance.