Universal descriptors of quasi transition states for small-data-driven asymmetric catalysis prediction in machine learning model

In enantioselectivity prediction models, the demand for numerous descriptors and extensive datasets poses a substantial challenge. Descriptor selection, fraught with uncertainty, compounds the issue, while amassing requisite data remains a daunting task. The introduction of descriptors derived from quasi transition states offers a promising avenue to alleviate this burden. However, the challenge of descriptor selection persists. Here, we report a small-data-driven model based on universal descriptors. Key differentiating properties between diastereomeric quasi transition states, encompassing single-point energies, frontier orbital energies, Cartesian forces, and charges of core atoms, are proposed as universal descriptors. The model's efficacy is validated through its application to the asymmetric aldol and Negishi reactions, utilizing 3 experimental variables, merely less than 9 universal descriptors, and fewer than 150 training samples. Moreover, a method is presented using Cartesian forces to rectify discrepancies between quasi transition states and true transition states. This strategy circumvents tedious large-scale descriptor exploration and screening, offering an efficient choice for small-data-driven enantioselectivity prediction.