Automated High-Throughput Virtual Screening of Catalysts via Templated Organic Reaction Pathway Construction: A Case Study on Suzuki–Miyaura Coupling Reaction

High-throughput virtual screening of catalysts in organic chemistry remains a grand challenge due to the prohibitive computational cost in exploring vast chemical spaces with quantum-chemical methods. Here, we develop the self-learning diffusion model coupled with potential energy surface exploration (SL-DM-PES) framework, which enables automated high-throughput virtual screening via templated organic reaction pathway construction. This self-learning (SL) framework integrates a general diffusion model (DM) for generating three-dimensional structures of reaction intermediates and transition states directly from two-dimensional molecular graphs, with generalized global neural network potential (GG-NN) calculations for rapid energy evaluation and structure optimization, namely the PES exploration. A high-order pair-reduced equivariant message passing neural network (HPNN-ET) is developed for DM, achieving high precision (RMSE ≤ 0.062 Å) and generality (up to 83 elements) for generating large complexes (up to 362 atoms). As a case study, we applied the SL-DM-PES framework to the Suzuki-Miyaura cross-coupling reaction, using one of the widely accepted mechanisms as the pathway template. Complete reaction profiles of 6883 diverse Pd-phosphine catalysts were generated within 286 GPU hours, costing only $80 overall (about $0.01 per catalyst). With the derived kinetic energy barriers, promising ligands can be predicted, and the prediction is supported by further experiments. SL-DM-PES not only demonstrates the high efficiency of HPNN-ET for complex organic reaction profile generation, but also provides a fast route for reaction screening from first-principles.