Deep learning for predictions of hydrolysis rates and conditional molecular design of esters

Background: The hydrolysis rate of an ester is essential for the choice of materials in sustainable and eco-friendly applications. Methods: In this work, the autoencoder (AE) model has been constructed to predict the hydrolysis rate by inputting SMILES and partial charges. Moreover, the conditional autoencoder (CAE) model has been developed to design chemical structures of esters that possess hydrolysis rates close to the desired value. Significant Findings: By implementing the SMILES enumeration technique and the attention mechanism, our AE model exhibits significantly better performance than SPARC based on the root mean square error. For six biodegradable esters that have no experimental rate constants, the predictions of our AE model are in agreement with those based on the activation energies calculated from Dmol³. To design an ester satisfying the desired conditions, our CAE model demonstrates its capability of providing the best candidates of esters and their rate constants based on structural similarity and the least difference of hydrolysis rates. The derived structures are similar to the desired structure and their rate constants are close to the targeted value.