Quantum Descriptor-Based Machine-Learning Modeling of Thermal Hazard of Cyclic Sulfamidates

Cyclic sulfamidates are commonly used building blocks
in organic
synthesis. Correct classification of their thermal criticality is
crucial for the safe use of these compounds in process development
and scale-up. In this study, building on our earlier work (Ferrari
et al., 2022), we focused on modeling the reaction enthalpy of a family
of 5-membered cyclic sulfamidates toward strong bases. The key challenge
for the modeling task was the sparse availability of measured reaction
enthalpies, with only 29 measurements available. To address this challenge,
we used descriptors based on the quantum-chemical properties of the
molecules, as they are more closely related to reaction enthalpies
than typical cheminformatics-based descriptors. This approach allowed
us to avoid relying solely on data-to-fit models and to focus instead
on modeling reaction enthalpies using chemistry-aware techniques,
which are more appropriate for small data sets. Three models were
constructed using the quantum-chemical descriptors: the first one
combining Partial Least Squares (PLS) regression with a Genetic Algorithm
(GA), the second one based on the Least Absolute Shrinkage and Selection
Operator (LASSO) method, and last, a Gaussian Process Regression (GPR)
model. The three models achieved coefficients of determination of
0.78, 0.67, and 0.74, respectively. Although the absolute prediction
error values were close to 100 J/g, it is noteworthy that all three
techniques provided similar results and accurately classified nearly
all compounds into their respective thermal criticality classes. This
highlights the methodology’s effectiveness in providing a reliable
framework for preliminary safety assessment and decision-making in
process development.