Physicochemically Informed Axial Chirality Descriptors Enable Accurate Prediction of Atropisomeric Stability

Atropisomers play a vital role in asymmetric synthesis, drug discovery, and the development of functional materials. However, the rational design of atropisomers is challenging due to the difficulty in predicting their configurational stability, which depends on the rotational barrier (ΔG^‡). Here, we introduce ACSD-GAT, a deep learning framework that addresses this issue. Our approach comprises a newly curated benchmark dataset of 1015 experimentally measured rotational barrier, along with a physicochemically informed axial chirality structure descriptor (ACSD) that explicitly quantifies both static and dynamic steric repulsion during rotation. By integrating the ACSD with a graph attention network (GAT), our model accurately predicts the rotational barrier, achieving an R² of 0.91 and a RMSE of 2.02 kcal mol^-1 on test datasets. The robustness and real-world applicability of the model are also demonstrated through rigorous validation with complex pharmaceuticals, molecular switches, and newly synthesized atropisomers.