Fig. 1. Overview of the proposed TSDiff.

a Illustration of the reaction graph (${\mathcal{G}}_{\mathrm{rxn}}$) and the denoising network. The molecular graphs of the reactants (R) and products (P), denoted ${\mathcal{G}}_R$ and ${\mathcal{G}}_P$ respectively, are constructed from the SMARTS representation of the reaction. Then, the condensed graph ${\mathcal{G}}_{\mathrm{rxn}}$ is formed using the node vectors ($\mathcal{V}$) and edge vectors ($\mathcal{E}$) obtained from them. The denoising network denoises a given geometry input based on ${\mathcal{G}}_{\mathrm{rxn}}$. b Transition state (TS) generation procedure of the proposed TSDiff. Starting from a randomly initialized geometry, the geometry is progressively refined by the denoising network until reaching a predicted TS geometry. All molecular geometries were plotted using PyMOL⁷¹.