Graph-based prediction of reaction barrier heights with on-the-fly prediction of transition states

The accurate prediction of reaction barrier heights is crucial for understanding chemical reactivity and guiding reaction design. Recent advances in machine learning (ML) models, particularly graph neural networks, have shown great promise in capturing complex chemical interactions. Here, directed message-passing neural networks (D-MPNNs) on graph overlays of the reactant and product graphs were shown to provide promising accuracies for reaction property prediction. They only rely on the change in molecular graphs as input and thus require no additional information during inference. However, the reaction barrier height intrinsically depends on the conformations of the reactants, transition state, and products, which are not taken into account in standard D-MPNNs. In this work, we present a hybrid approach where we combine the power of D-MPNNs predicting barrier heights with generative models predicting transition state geometries on-the-fly, only given the reaction graph. The resulting model thus only requires two-dimensional graph information as input, while internally leveraging three-dimensional information to increase accuracy. We furthermore evaluate the influence of additional physical features on D-MPNN models of reaction barrier heights, where we find that additional features only marginally enhance predictive accuracy and are especially helpful for small datasets. In contrast, our hybrid graph/coordinate approach reduces the error of barrier height predictions for all investigated datasets.