Testing the Limits of Latent Diffusion for Transition-State Generation
ORAL
Abstract
Generative models such as denoising diffusion and flow matching have shown strong performance in molecular and materials generation [1–4]. Some of these architectures perform the generation in a latent space using Variational Autoencoders (VAE). VAEs provide compact representation that mitigates the scalability issues of high-dimensional atomic systems, and the architecture is proposed as a step toward a foundation model for generative chemistry. To that extent, it is of interest to assess whether such latent-space architectures also perform well in low-probability transition regions that govern chemical reactivity.
We therefore compare the VAE-based latent flow used in the recent All-atom Diffusion Transformer (ADiT) [1] with the current state-of-the-art direct conditional flow-matching model, TS-GEN [5]. ADiT employs a VAE with a Transformer-based latent flow to generate molecules and crystals, while TS-GEN maps Gaussian priors to transition-state geometries conditioned on reactant and product structures. Unlike ADiT, which performs flow matching in a learned latent space, TS-GEN applies flow matching directly in atomic coordinate space using an equivariant transformer backbone, enabling high geometric fidelity for transition-state prediction. We benchmark these approaches on two datasets, Transition1x and RGD1-CHNO, to evaluate their accuracy and transferability and suggest improvements.
[1] C. K. Joshi et al., All-atom Diffusion Transformers: Unified generative modelling of molecules and materials, ICML 2025.
[2] Y. Lipman et al., Flow Matching for Generative Modeling, ICLR 2023.
[3] W. Peebles and S. Xie, Scalable Diffusion Models with Transformers, ICCV 2023.
[4] R. Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models, CVPR 2022.
[5] P. Tuo et al., Accurate Generation of Chemical Reaction Transition States by Conditional Flow Matching, arXiv:2507.10530 (2025).
We therefore compare the VAE-based latent flow used in the recent All-atom Diffusion Transformer (ADiT) [1] with the current state-of-the-art direct conditional flow-matching model, TS-GEN [5]. ADiT employs a VAE with a Transformer-based latent flow to generate molecules and crystals, while TS-GEN maps Gaussian priors to transition-state geometries conditioned on reactant and product structures. Unlike ADiT, which performs flow matching in a learned latent space, TS-GEN applies flow matching directly in atomic coordinate space using an equivariant transformer backbone, enabling high geometric fidelity for transition-state prediction. We benchmark these approaches on two datasets, Transition1x and RGD1-CHNO, to evaluate their accuracy and transferability and suggest improvements.
[1] C. K. Joshi et al., All-atom Diffusion Transformers: Unified generative modelling of molecules and materials, ICML 2025.
[2] Y. Lipman et al., Flow Matching for Generative Modeling, ICLR 2023.
[3] W. Peebles and S. Xie, Scalable Diffusion Models with Transformers, ICCV 2023.
[4] R. Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models, CVPR 2022.
[5] P. Tuo et al., Accurate Generation of Chemical Reaction Transition States by Conditional Flow Matching, arXiv:2507.10530 (2025).
–
Presenters
-
Ulrik Unneberg
- Harvard University