Multi-modal Homogeneous Chemical Reaction Performance Prediction with Graph and Chemical Language Information

Abstract

Accurate prediction for chemical reaction performance offers optimal direction for synthetic development. To this end, we present a novel multi-modal model called MMHRP-GCL to achieve the prediction of homogeneous chemical reaction yield, enantioselectivity, and activation energy by fusing the information from the text and graph modalities, requiring only 8 simple descriptors and Reaction SMILES obtained without high-cost DFT computation, and capable of managing reactions involving a fluctuating number of molecules. Experimental results on 4 datasets show that MMHRP-GCL outperforms at least 7 generalized SOTA methods. Ablation study confirms the critical roles of the complementation of graph and text modalities, as well as the significance of modality alignment and atomic features in prediction. Albeit there is still room for improvement in the interpretation of atomic relationships, the model has a remarkable ability to identify important atoms. A statistically interpretable study of the feature importance and a test on challenging dataset further demonstrates the utility and potential of the model. As a high-accuracy, low-cost, interpretable, and general multi-modal model, MMHRP-GCL provides valuable guidance on the design of forward predictors for homogeneous catalytic reactions.

Keywords

Reaction prediction / Deep learning / Multimodal model / Interpretable / Computational chemistry / Molecular modeling / Homogeneous catalysis

RIGHTS & PERMISSIONS

2025 SIOC, CAS, Shanghai, & WILEY-VCH GmbH.